allocate_convention_ids
ChatGPTGenerate fresh pattern and output dimension IDs for a new convention or newly added pattern/dimension. Use these values verbatim in prepare_create_convention or prepare_update_convention; never replace IDs of existing dimensions. This allocates identifiers only and changes no workspace data.
apply_field_mutation
ChatGPTWrite the fields the user approved from one mutate_fields proposal. The plan is rebuilt and checked against each field's current definition; a field changed since preparing is reported as stale and is not written. A plan applies at most once and expires after an hour. Each field is written once and never retried: a lost answer is read back, and anything still unconfirmed is reported as unknown. Unknown fields need a fresh look at their current definition before any further change.
check_portal_dashboards
ChatGPTSays what add_dashboards_to_portal would do, without writing: each dashboard is read from dashboards-service in the call's workspace, with its title. addable are dashboards in the workspace and not yet in the portal, already_in_portal those the portal has, not_found ids that are not dashboards in the workspace. not_found for the portal itself when the workspace has no such portal. Needs get_account. Use it to show the user what they approve.
copy_fields
ChatGPTCopy up to 25 custom fields from source_workspace_id into this workspace, keeping their ids (Fields API POST /v2/{workspace}/field-imports). dry_run: true writes nothing and returns checks; dry_run: false creates the fields and returns them with the same checks. Each check names missing_dependencies (fields neither workspace will have, such as a connector this workspace lacks: the copy returns no data for them), differing_dependencies (fields defined differently here), source_specific_rules (rules that match a data source this workspace does not have, so they never match) and imported_dependencies (the custom fields this copy adds). Fields with findings are still copied. include_dependencies: true also copies the custom fields the requested ones use that this workspace lacks (requested: false in their checks); a real copy is invalid, writing nothing, when that makes more than 25 fields in all. A real copy writes nothing when a field is missing from the source (not_found) or already exists here (already_exists); existing fields are never overwritten. A field created here while the copy runs is also already_exists, and details then names the fields copied before the conflict. Needs get_account and edit_fields on both workspaces, which must belong to the same subscription. workspace_id is the workspace that receives the copies. Call it with dry_run: true first, show the user every finding in checks, and copy with dry_run: false only after the user confirms.
delete_data_source
ChatGPTSoft-delete one existing source by exact UUID. Inspect first; present workspace, source name/connector/state and impact, then obtain separate explicit approval of THIS deletion. It does not require a connection or template. Depository marks it DELETED, sets a 31-day record expiry and attempts an ARCHIVED read blocker; it does not itself purge already downloaded data and blocker failure may not prevent deletion. Pass inspected name and state as stale preflights, not an atomic lock. DELETE once; if uncertain inspect, never retry automatically. Restoration is a separate action requiring its own approval.
execute_brand_change
ChatGPTMake the brand change a prepare_create_brand, prepare_update_brand or prepare_delete_brand proposal describes. Call it only after the user approved that change. A proposal executes at most once and expires after 30 minutes; prepare again after that. An update applies the approved fields to the brand as it is now. One write, never retried: after an unknown outcome, read the brands back with list_brands before preparing it again.
execute_budget_change
ChatGPTMake the budget change a prepare_update_budget, prepare_archive_budget or prepare_restore_budget proposal describes, exactly as proposed. Call it only after the user approved that change. A proposal executes at most once and expires after 30 minutes; prepare again after that. An update answers stale, writing nothing, when the budget was edited after it was proposed: prepare it again. One write, never retried: after an unknown outcome, read the budget back with list_budgets or get_budget before preparing it again.
execute_convention_change
ChatGPTExecute exactly the approved prepare_create_convention, prepare_update_convention, prepare_archive_convention or prepare_restore_convention proposal. Call it only after the user approved that change. A proposal executes at most once and expires after 30 minutes; prepare again after that. Update, archive and restore answer stale without writing if the convention changed; prepare again and obtain new approval. A change can affect multiple linked dimensions and is not all-or-nothing. When the result has an app_url, give it to the user: it opens the changed item in Funnel. One write, never retried: after an unknown outcome, read back with list_conventions, list_archived_conventions or get_convention before preparing it again.
execute_create_data_source
ChatGPTCreate the data source a prepare_create_data_source proposal describes, exactly as proposed. Call it only after the user approved that change. The setup is checked again first; if anything changed it is reapproval_required and nothing is created. A proposal executes at most once and expires after 30 minutes. created is a source record, not proof downloads work. One create, never retried: after an unknown outcome, look for the source with search_fields category=source before preparing it again.
execute_create_workspace
ChatGPTCreate the workspace a prepare_create_workspace proposal describes, exactly as proposed, with the user as its editor. Call it only after the user approved that change. A proposal executes at most once and expires after 30 minutes; prepare again after that. not_allowed created nothing and says why. One create, never retried: after an unknown outcome, ask the user to check the subscription's workspaces in Funnel before preparing it again.
execute_dashboard_change
ChatGPTMake the dashboard change a prepare_create_dashboard, prepare_append_dashboard_cards or prepare_update_dashboard_cards proposal describes, exactly as proposed. Call it only after the user approved that change. A proposal executes at most once and expires after 30 minutes; prepare again after that. An update answers stale, changing nothing, when the dashboard changed after it was proposed: read it again and prepare again. The result's dashboard.digest is the based_on_digest of a next prepare_update_dashboard_cards on that dashboard; give the user dashboard.app_url, which opens it in Funnel. One write, never retried: after an unknown outcome, read the dashboard back with list_dashboards or get_dashboard before preparing it again.
execute_dashboard_share_link_change
ChatGPTRun the dashboard share link change a prepare_invalidate_dashboard_share_link or prepare_create_dashboard_share_link proposal describes. Call it only after the user approved that change. A proposal executes at most once and expires after 30 minutes; prepare again after that. already_invalidated means the link had stopped working already: nothing more to do. created answers the new link and its url. One write, never retried: after an unknown outcome, read list_dashboards' share_links before preparing it again.
execute_export_change
ChatGPTMake the export change a prepare_create_export, prepare_create_view, prepare_set_export_enabled or prepare_run_export proposal describes, exactly as proposed. Call it only after the user approved that change. A proposal executes at most once and expires after 30 minutes; prepare again after that. When the result has an app_url, give it to the user: it opens the changed item in Funnel. One write, never retried: after an unknown outcome, read the export back with list_exports or get_export before preparing it again; for a view, ask the user to check the views in Funnel.
execute_field_change
ChatGPTDelete the custom field a prepare_delete_field proposal describes, exactly as proposed. Call it only after the user approved that deletion. A proposal executes at most once and expires after 30 minutes; prepare again after that. status: deleted, or stale (the field was edited after it was proposed; nothing was deleted: prepare it again and ask again). One write, never retried: after an unknown outcome, look the field up with search_fields before preparing it again.
execute_portal_change
ChatGPTMake the portal change a prepare_create_portal, prepare_update_portal, prepare_delete_portal, prepare_add_dashboards_to_portal, prepare_remove_dashboards_from_portal or prepare_set_portal_dashboard_groups proposal describes. Call it only after the user approved that change. A proposal executes at most once and expires after 30 minutes; prepare again after that. Dashboards are checked again: adding skips any now in the portal or gone, removing skips any the portal no longer has. One write, never retried: after an unknown outcome, read the portal back with get_portal or list_portal_dashboards before preparing it again.
execute_rename_workspace
ChatGPTRename the workspace as a prepare_rename_workspace proposal describes, exactly as proposed. Call it only after the user approved that change. A proposal executes at most once and expires after 30 minutes; prepare again after that. not_allowed renamed nothing and says why. After an unknown outcome, prepare the same name again: unchanged means the rename went through.
execute_scheduled_task_change
ChatGPTMake the scheduled task change a prepare_create_scheduled_task, prepare_update_scheduled_task or prepare_delete_scheduled_task proposal describes, exactly as proposed. Call it only after the user approved that change. A proposal executes at most once and expires after 30 minutes; prepare again after that. An update or delete answers stale, changing nothing, when the task changed after it was proposed: tell the user and prepare it again. One write, never retried: after an unknown outcome, read the tasks back with list_scheduled_tasks before preparing it again.
execute_update_workspace_settings
ChatGPTChange the workspace's settings as a prepare_update_workspace_settings proposal describes, exactly as proposed; settings it does not change are kept. Call it only after the user approved that change. A proposal executes at most once and expires after 30 minutes; prepare again after that. not_allowed changed nothing and says why. After an unknown outcome, prepare the same change again: unchanged means it went through.
execute_workspace_user_change
ChatGPTMake the user changes a prepare_set_workspace_users proposal describes, exactly as proposed. Call it only after the user approved that change. A proposal executes at most once and expires after 30 minutes; prepare again after that. completed gives each email's outcome. Changes are never retried: for not_sent emails, prepare them again; for uncertain ones, ask the user to check the workspace's users in Funnel first.
execute_workspace_user_removal
ChatGPTRemove the members a prepare_remove_workspace_users proposal lists, exactly as proposed. Call it only after the user approved that change. A proposal executes at most once and expires after 30 minutes; prepare again after that. not_allowed removed no one and says why. completed gives each email's outcome. Removals are never retried: for not_sent emails, prepare them again; for uncertain ones, ask the user to check the workspace's users in Funnel first.
forecasting___create_forecast_scenario
ChatGPTForecasting (what-if on the deployed MMM model): save and run a forecast scenario, the same as the Forecast page's "Fixed scenario". It is saved as a fixed media plan, so it is listed and compared with the other plans. Use for questions like: - "Save a scenario with Search +20% and Display -10% for the next 90 days." - "Create a what-if scenario I can come back to." Not for: - A quick unsaved answer -> run_forecast_what_if - Optimised budget allocation or the budget needed for a target -> create_media_plan - Changing a saved scenario -> update_forecast_scenario Before calling: account_id and partition_id; the partition needs a successfully deployed MMM model. The tool reads the deployed graph and the channel totals itself. Defaults, as in the UI: - Window: tomorrow for 90 days (start_date and end_date can be changed; end at most 730 days from today). The reference window is the same dates a year earlier. - Every channel (media budgets and impressions & clicks) starts at 0%, its total for the window. channel_changes_pct moves a channel by -100..100 percent in steps of 5, like the UI slider. A channel with no spend in the window cannot be changed. - Every non-binary control variable is {"type": "use_reference"}. Override per name with {"type": "fixed", "value": v} or {"type": "use_provided_data"}. Unknown channel or control variable names are rejected with the valid names. Always call first without confirmed: the preview's proposed_change lists every channel (window total, change and resulting value) and every control variable, plus warnings. Show them to the user and offer to change any before calling again with confirmed=true and the same config. Example config: {"name": "Search push", "channel_changes_pct": {"Search - Branded": 20, "Display": -10}} Returns: the saved plan (its id is the media_plan_id) and the run that was started, like the UI's "Create scenario". Poll get_media_plan_run_status, then read results with the media plan result tools.
forecasting___update_forecast_scenario
ChatGPTForecasting (what-if on the deployed MMM model): change a saved forecast scenario (a fixed media plan) and rerun it, like the UI's "Save & Retrigger". Use for questions like: - "Push Search to +30% in my saved scenario." - "Move the scenario to start next month." Not for: - Optimiser media plans -> update_media_plan - A new scenario -> create_forecast_scenario Before calling: media_plan_id of a scenario_type=fixed plan from list_media_plans. Send only what changes: - channel_changes_pct: percent from the channel's current total for the window (-100..100, step 5). - control_variables: per-name overrides, as in create_forecast_scenario. - start_date / end_date: changing the window resets every channel to its new window total, as the UI does; send channel_changes_pct in the same call to change them again. Other channels keep their saved values; the preview shows them as a percent of the current window total. A name-only change is saved without a rerun. Always call first without confirmed: proposed_change holds every resolved value, a changes list (from -> to) and will_rerun. Show them to the user before calling again with confirmed=true and the same config. Returns: the changes, the updated plan, and the run that was started (if any).
funnel___get_dimension_values
ChatGPTReturns actual values for a dimension field within a date range — e.g., campaign names, country codes, ad group names, platform names. Useful for previewing what filter values exist before building a query, or for answering "what campaigns ran in Q1?" directly. Inputs: workspace ID, dimension field ID (from search_fields or load_workspace), date range. For platform-specific dimensions on workspaces with multiple accounts on the same platform, a source ID is needed to disambiguate.
funnel___get_workspace_context
ChatGPTRetrieves workspace context — administrator-authored conventions, business rules, and domain terminology — to improve analysis accuracy and terminology alignment. Call once per workspace after load_workspace, before analyzing data. The context is returned as markdown in content[0].text and is auto-injected into the model's understanding. Cached per session (5-minute TTL) — repeated calls return instantly. Falls back gracefully if context is unavailable.
funnel___list_workspaces
ChatGPTLists Funnel workspaces accessible to the current user, grouped by subscription. Each subscription represents an organisation in Funnel. Within each subscription, workspaces represent individual advertisers or business units with their own connected ad platforms (Google Ads, Meta, LinkedIn, etc.). A user may have access to workspaces across multiple subscriptions. The returned workspace IDs are used by all other tools. Invoke when the target workspace is unknown, or to discover which workspaces are available for analysis.
funnel___load_workspace
ChatGPTLoads a workspace's available data: ad platforms, dimensions (campaign name, country, ad group, etc.), measures (spend, clicks, impressions), and metrics (CPC, CTR, ROAS, etc.). Returns a summary of connected platforms with field counts, source health status, and which cross-platform metrics are available. Cached per session — repeated calls for the same workspace return instantly. Required once per workspace before querying data or searching for fields. Input: workspace ID from list_workspaces. After loading, call get_workspace_context once for this workspace to retrieve administrator-defined conventions, business rules, and domain terminology that improve analysis accuracy and terminology alignment.
funnel___prepare_data
ChatGPTLoads raw, row-level data without aggregation — every individual data point as reported by the ad platform. Intended for scenarios where the same data needs to be sliced multiple ways (use query_data mode="aggregate" afterward), or when individual-row inspection is needed. Unlike query_data mode="query" which aggregates at load time (e.g., summing daily spend by campaign), this preserves the raw grain. Returns available ratio metrics that can be computed on the loaded data. Inputs: workspace ID, unique table name, dimensions, measures, date range. Platform-specific dimensions need a source ID. Surfaces a source selection prompt when multiple sources exist for the same platform.
funnel___query_data
ChatGPTQueries advertising data from connected platforms. Returns tabular results into a named session table for further analysis. 1. Pick a mode: - query — fetch fresh data by dimensions and measures for a date range - compare_channels — compare metrics across all platforms (cross-platform fields ONLY) - aggregate — re-slice a previously loaded table at different grain (no new fetch) - compute_metric — calculate a ratio (CPC, CTR, ROAS) on a loaded table - restore — reload data evicted from session memory 2. SCOPING RULES (critical — most errors happen here): - Each field is scoped to cross_platform or a specific platform. - Platform-specific fields require source_id (auto-resolves for single-source platforms). - All-cross-platform queries must omit source_id (returns workspace-wide totals). - All platform-specific fields in a query must belong to the same platform. 3. Required fields per mode: - query: workspace_id, result_table_name, dimensions (use [] for totals), measures, date_start, date_end - compare_channels: workspace_id, result_table_name, measures, date_start, date_end (optional: metrics for ratio computation) - aggregate: workspace_id, source_table_name - compute_metric: workspace_id, source_table_name, ratio_id - restore: workspace_id, evicted_table_name 4. Source disambiguation: When multiple sources exist and none is specified, the response returns source_selection_required with available_sources grouped by account. Platform-specific queries accept source_id to resolve. Cross-platform queries accept source_overrides with one source_id per group (a platform may have multiple groups if it has multiple accounts). Dates: YYYY-MM-DD. Percentages as decimals (0.021 = 2.1%). Default limit 1000 rows. Examples: Cross-platform totals: {"mode":"query","workspace_id":"<id>","result_table_name":"totals","dimensions":[],"measures":["common-cost","common-impressions"],"date_start":"2025-01-01","date_end":"2025-03-31"} Single-source with platform fields: {"mode":"query","workspace_id":"<id>","result_table_name":"fb_campaigns","dimensions":["facebookads-campaign_name"],"measures":["common-cost","common-clicks"],"source_id":"<source_id>","date_start":"2025-01-01","date_end":"2025-03-31"} Daily breakdown (re-aggregate): {"mode":"aggregate","source_table_name":"fb_campaigns","group_by":["facebookads-campaign_name","date"],"measures":["common-cost"],"date_granularity":"day"}
funnel___search_fields
ChatGPTFinds available dimensions, measures, and metrics in a workspace. Three modes: semantic (preferred) — AI-powered field discovery. Classifies your intent into structural filters (category, platform), narrows the catalog, then selects the most relevant fields by reading their labels and descriptions. Handles synonyms, acronyms, and cross-platform vs platform-specific disambiguation automatically. Ask natural language questions like "how much did I spend on Facebook", "hubspot demo conversions", "cross-platform performance metrics", "pipeline velocity". browse — Navigate fields by platform, category (dimension/measure/metric), or group. Use for structured exploration or listing data sources with their IDs. search — Keyword text matching. Use only when you know the exact field name or label. Prefer semantic mode for discovery. Use browse to drill into a known group. Use search only for exact lookups.
get_budget
ChatGPTReads one active budget of the workspace (plan-api GET /{workspace}/budgets/{id}): its dates, its targets with their totals, units and number of periods, its rollover and pacing settings, and its page in Plan. Spend against it comes from get_data with budget_id. not_found when the workspace has no such active budget. Needs get_account.
get_convention
ChatGPTRead one active Plan convention by id from list_conventions, including its patterns, field mappings, output-field settings and validation rules. Returns not_found for a missing or archived convention. Needs get_account permission. Not available in headless runs.
get_creative_images
ChatGPTLook at the creative images of the top ads yourself: returns each creative's first image as an image you can view, with its label and metric values. Use it when the user asks what the creatives look like, or to describe, compare or explain creatives by their visuals; use visualize_creatives to show the grid to the user. Takes the same arguments as visualize_creatives except title, for up to 6 creatives (limit 1 to 6, default 4). Videos and images over 1 MB are named but not returned.
get_dashboard
ChatGPTReads one dashboard of the workspace (dashboards-service GET /accounts/{workspace}/dashboards/{id}): its title, its config (data tables, cards, layout, filters, date range), its feature flags, when it was last changed and the digest of its config and title. Read it before proposing a change to it, and copy what does not change. Counts a dashboard view. not_found when the workspace has no such dashboard. Needs get_account. To reproduce a card's numbers, call query_data with the metrics and dimensions the card's data names (such as metricId, dimensionId, metrics, dimensions), over the card's data.overrideDateRange or else the dashboard's defaultDateRange (an override may set only compareTo; resolve a presetRangeId to dates; query the compareTo period as a second date range), and translate into query_data filters the query.where of the card's data table (dataTableId), each of the dashboard's globalFilters that has a condition, and the card's data.whereAllTrue. Use the dashboard's currency when set. Tell the user about any setting you could not translate, and share app_url so they can open the dashboard in Funnel.
get_dimension_values
ChatGPTReturns actual values for a dimension field within a date range. scope controls the Data Source boundary: inferred resolves from the dimension, workspace accepts cross-platform dimensions, and data_source enforces one immutable ID. Omitted scope preserves pre-typed compatibility behavior during migration. Deprecated source_id remains accepted but cannot be combined with scope.
get_explorer_view
ChatGPTRead one saved Data Explorer view by id from list_explorer_views: its Funnel field ids (fields, split into dimensions and metrics when the view's table says so), date range (preset such as last30days, or start/end, plus compare_to), currency, sort, and filters. where (dimension filters) and having (metric filters) are Data Explorer filter expressions such as {"sourceName":{"=icontains":"global"}}. The view holds no data: query its fields and date range to get the numbers. Returns not_found for a missing or archived view. Needs get_account permission. Not available in headless runs. To get the view's numbers, call query_data with its dimensions and metrics (classify fields with search_fields when they are not split), the date range, its currency, and the where conditions translated into query_data filters. Tell the user about any condition you could not translate, and share app_url so they can open the view in Funnel.
get_export
ChatGPTLoads one export's or view's configuration (any id from list_exports) and returns a ready-to-use query template (dimensions, measures, metrics, date range, currency, filters, scope) for reproducing the exported data with query_data. A currency of * means the export converts nothing: pass currency * (native) rather than omitting it, which would convert to the workspace default. Check unsupported and unknown_fields: when non-empty the reproduction is approximate. A view stores no dates (the BI tool picks them), so its template has none: ask the user for the range.
get_export_run
ChatGPTLoads one export run (a run_id from list_export_runs): its status and counts, every part (date period) it produced with row and column counts, file size and a download link, and the failed or warning steps with their error code and message. A download link works for about an hour without signing in, so share it only with the user who asked. Needs get_account.
get_flexpoint_usage
ChatGPTThe Flexpoints the workspace's subscription uses, its capacity and what is available, as the subscription's Flexpoints usage page shows them (credits-api GET /usage/subscription/{subscription}/summary). Use available for how many are left. Needs get_subscription_usage on the subscription. A subscription without a Flexpoints plan is not_enabled.
get_flexpoint_usage_by_connector
ChatGPTFlexpoints per connector and destination type across the workspace's subscription, highest first, as the Resources table of the subscription's Flexpoints usage page (credits-api GET /usage/subscription/{subscription}/resourceSummary). Paused resources use no Flexpoints, as on the page, and are left out unless status asks for them. Needs get_subscription_detailed_usage on the subscription.
get_flexpoint_usage_by_workspace
ChatGPTThe Flexpoints attributed to each workspace of the workspace's subscription, highest first, with the subscription's pool, as the Workspaces tab of its Flexpoints usage page (credits-api GET /usage/subscription/{subscription}/workspace-summary). Workspaces share the pool, so only the pool has capacity. Needs get_subscription_detailed_usage on the subscription. A subscription without a Flexpoints plan is not_enabled.
get_portal
ChatGPTReads one portal of the workspace by id (portal-service GET /accounts/{workspace}/portals/{id}); not_found when the workspace has no such portal. Needs get_account.
get_portal_dashboard_groups
ChatGPTReads how a portal's dashboards are grouped and ordered for its viewers (portal-service GET /accounts/{workspace}/portals/{id}/dashboard-sort-config), with each dashboard's title as the portal shows it. Groups are in display order; the group with name: null holds the ungrouped dashboards. Dashboards the portal has that no group lists are appended to it, and ids the portal no longer has are dropped, as the Portals page does. not_found when the workspace has no such portal. Needs get_account.
get_scheduled_task_runs
ChatGPTGet the recent run history for a scheduled task by id: whether each fire completed or failed, when, and a short preview of the output (the full result is in the linked conversation). Use to answer 'did my scheduled report run?' or to debug a failing task. not_found when the workspace has no such task. Run output was produced with the owner's access, so this needs get_account and edit_account (workspace admins); other members can list tasks but not read their runs.
get_workspace_context
ChatGPTRetrieves workspace context — administrator-authored conventions, business rules, and domain terminology — to improve analysis accuracy and terminology alignment. Call once per workspace after load_workspace, before analyzing data. The context is returned as markdown in both content[0].text and the structured workspace_context field so clients with either result-channel behavior can inject it into the model's understanding. Cached per session (5-minute TTL) — repeated calls return instantly. Falls back gracefully if context is unavailable.
incrementality___create_incrementality_test_from_candidate
ChatGPTIncrementality (geo lift testing): register a geo lift test from one candidate of a finished design, like the UI's "Create test" on a candidate. Use for questions like: - "Create a test from the top candidate of my design." - "Set up candidate 3 as a test starting on 1 March." Not for: - A test that does not come from a design -> create_incrementality_test - Comparing candidates first -> get_incrementality_design_candidates, get_incrementality_candidate_power Before calling: design_id with status SUCCEEDED; candidate_rank from get_incrementality_design_candidates. The test is linked to the design and takes its model and metric; granularity is daily, as in the UI. Every other field defaults from the design and candidate as the UI fills its form, and can be overridden. Always call first without confirmed=true and present the preview (regions, dates, planned spend, test type). Returns: the created test with test_id.
inspect_data_source
ChatGPTRead the management state of one exact Data Source UUID from search_fields category=source or list_deleted_data_sources. Resolves its private Depository key within the authorized workspace; a missing or inaccessible source returns not_found. Never changes the source.
list_archived_budgets
ChatGPTLists the workspace's archived Plan budgets (plan-api GET /{workspace}/budgets/archive), sorted by name; updated_at is when it was archived or last changed. Use it to find the id prepare_restore_budget takes. Archived budgets have no data source, so get_data has no series for them. Needs get_account on the workspace.
list_archived_conventions
ChatGPTList archived Plan conventions in a workspace, sorted by name. These are not active and cannot be fetched with get_convention. Needs get_account permission. Not available in headless runs.
list_brands
ChatGPTLists the portal brands of the subscription that owns the workspace (portal-service GET /subscriptions/{subscription}/brands), by name. Needs get_account on the workspace and get_subscription on its subscription.
list_budgets
ChatGPTLists the workspace's configured Plan budgets (plan-api GET /{workspace}/budgets/v2; archived budgets excluded), sorted by name. Match the user's wording against name and description, then pass the id to get_data as budget_id. Needs get_account on the workspace.
list_connector_templates
ChatGPTList curated Funnel template candidates for one connector in the selected workspace without a setup snapshot or an existing source. Use the connector_id from list_connectors or source inspection. This read does not establish whether an existing source may switch to a candidate; Depository validates reconfiguration. No definitions or credentials are returned.
list_connectors
ChatGPTList the bounded Funnel connectors available for data source setup in the selected workspace. Pass a returned connector_id to start_data_source_setup or start_data_source_authentication.
list_conventions
ChatGPTList active Plan conventions in a workspace, sorted by name. Returns names, levels, input fields and counts; use get_convention for patterns and output fields. Needs get_account permission. Not available in headless runs.
list_dashboards
ChatGPTLists the workspace's dashboards by title, with their ids (dashboards-service GET /accounts/{workspace}/dashboards), up to 1000. Reading the list counts no dashboard view. Needs get_account on the workspace.
list_data_source_templates
ChatGPTList eligible curated Funnel templates for the remote account in an immutable setup snapshot. This does not create or modify a data source.
list_data_sources
ChatGPTList active (not deleted) sources in the selected workspace to find the exact public UUID for inspect_data_source and the source management tools. Unlike search_fields, it includes sources excluded from Funnel's data. Pass search (case-insensitive name substring) to find a named source and connector_id to narrow to one connector; page with offset if has_more (offset stops at 9,900; beyond that, narrow with search or connector_id). No private Depository keys or credentials are returned. Identify one source by customer-visible name and connector, then inspect its UUID; if several match, ask the user. Do not change a source from an ambiguous or incomplete result.
list_deleted_data_sources
ChatGPTList soft-deleted sources in the selected workspace to discover an exact public UUID for restore_data_source. Uses a workspace-scoped DELETED-state search; no private Depository keys or credentials are returned. Page with offset/limit if has_more; identify one source by customer-visible name and connector, then inspect its UUID. Do not guess or restore a source from an ambiguous or incomplete result.
list_explorer_views
ChatGPTList the workspace's saved Data Explorer views, sorted by name: id, name, last update and a link to open the view in Funnel. Archived views are left out. Use get_explorer_view for a view's fields, date range and filters. Needs get_account permission. Not available in headless runs.
list_export_connections
ChatGPTLists the caller's own connections that can write to a destination type (Connection API GET /v3/connection): Google connections that granted the destination's scopes (Sheets for gsheets; BigQuery and Cloud Storage for bigquery; Cloud Storage for gcs), and Snowflake ones (a user with a programmatic access token or a key pair) for snowflake. Pass one id as the destination's connection_id in prepare_create_export. When connections is empty, the user must connect an account first: give them connect_url, a Funnel page where they sign in (the credentials are entered there, never in chat), then list again once they say they are done. A null connect_url means the account must be connected on the new-export page in Funnel (app_url). Never changes anything.
list_export_destinations
ChatGPTLists the destinations Funnel data can be exported to, and how each is created: from chat (create: chat, with tool) or only in Funnel (create: app, give the user app_url). Whether the workspace's plan allows a destination is checked when it is created. Never changes anything.
list_export_runs
ChatGPTLists one export's recent runs, newest first: when each started, its status (Success, Failed, Pending, Stopped), the error code and message of a failed run, and how many of its parts (date periods) succeeded, warned or failed. Use to answer 'did my export run?' or 'why did it fail?', then get_export_run for one run's parts, file sizes and download links. Pass next_cursor back as cursor for older runs. A view, or an export that never ran, has runs: []. Needs get_account.
list_exports
ChatGPTLists the data exports configured in a workspace — scheduled pushes of Funnel data to Google Sheets, Excel, BigQuery, Snowflake, S3 and similar destinations — and its views, which a BI tool (Looker Studio, Tableau, Power BI) reads from Funnel (kind). Match the user's wording against name and remote_target, then call get_export with the item's id and kind. Only an export (kind: export) can be turned on or off or run.
list_guides
ChatGPTList curated Funnel guides when guide_id is omitted. With guide_id=field-rules, read the custom field rule and formula authoring guide. Read-only; mutate_fields may not be available in every environment.
list_portal_dashboards
ChatGPTLists the dashboards in one portal of the workspace, by the title the portal shows (portal-service GET /accounts/{workspace}/portals/{id}/items), with the portal itself. not_found when the workspace has no such portal. Needs get_account.
list_portal_users
ChatGPTLists the users who can open one portal of the workspace (portal-service GET /accounts/{workspace}/portals/{id}/users), newest first, up to 1000, with the portal itself. Each user's email, display name and workspace roles come from auth-service's workspace users (GET /accounts/{workspace}/users); they are null for a user no longer in the workspace, except that a user id that is an email is its own email. not_found when the workspace has no such portal. Needs get_account.
list_portals
ChatGPTLists the workspace's client portals with their brand, dashboard count and user count (portal-service GET /accounts/{workspace}/portals), newest first, up to 500. Needs get_account on the workspace.
list_scheduled_tasks
ChatGPTList this workspace's scheduled tasks (recurring prompts that run automatically as their owner). Use to answer 'what do I have scheduled?', to find a task's id before editing/deleting it, or to confirm a task exists. Returns each task's id, name, prompt, schedule expression, timezone, status (enabled/disabled) and last run id. A disabled task may carry paused_reason (e.g. owner_lost_access = the owner no longer has access; a workspace member must resume it). Needs get_account on the workspace.
list_workspace_users
ChatGPTLists the members of this workspace, as Funnel's Users page does (auth-service GET /api/v1/accounts/{workspace}/users): each member's user id, email, display name and workspace roles, sorted by email, up to 1000 (truncated when there are more). Subscription owners and admins also have editor access to every workspace of the subscription; they are listed only when they are members. app_url is the workspace's users page. Needs get_account on the workspace.
list_workspaces
ChatGPTLists Funnel workspaces accessible to the current user, grouped by subscription. Each subscription represents an organisation in Funnel. Within each subscription, workspaces represent individual advertisers or business units with their own connected ad platforms (Google Ads, Meta, LinkedIn, etc.). A user may have access to workspaces across multiple subscriptions. The returned workspace IDs are used by all other tools. Invoke when the target workspace is unknown, or to discover which workspaces are available for analysis.
load_workspace
ChatGPTLoads a workspace's available data: ad platforms, dimensions (campaign name, country, ad group, etc.), measures (spend, clicks, impressions), and metrics (CPC, CTR, ROAS, etc.). Returns a summary of connected platforms with field counts, source health status, and which cross-platform metrics are available. Cached per session — repeated calls for the same workspace return instantly. Required once per workspace before querying data or searching for fields. Input: workspace ID from list_workspaces. After loading, call get_workspace_context once for this workspace to retrieve administrator-defined conventions, business rules, and domain terminology that improve analysis accuracy and terminology alignment.
measurement_check_incrementality_data_readiness
ChatGPTIncrementality (geo lift testing): run a pre-flight check on the data before designing a geo test and return GO, CAUTION or STOP. Use for questions like: - "Is my data good enough for a geo test?" - "Can I run an incrementality test on revenue?" Not for: - Exploring the data -> get_incrementality_geo_data - Only the correlation views -> get_incrementality_region_correlation Before calling: metric_name from list_incrementality_metrics. Run this before create_incrementality_test_design on a new metric or partition; do not judge feasibility from get_incrementality_geo_data alone. Returns: verdict, reasons[], and data_quality, comovement and factor_structure summaries.
measurement_compare_media_plans
ChatGPTMedia plans (budget optimiser and saved what-if scenarios): compare every media plan in this partition by outcome and rank them. Use for questions like: - "Compare my media plans." - "Which plan performs best?" Not for: - One plan's detailed results -> get_media_plan_projected_results, get_media_plan_pacing Before calling: account_id and partition_id from list_accounts_and_partitions. Returns: {count, plans[]} ranked by ROAS, highest first. Each plan carries its config fields and metrics; metrics_source says whether they came from the optimiser projection, the reports optimal series or actuals. Plans without results have metrics: null; if their results could not be loaded, metrics_error says why.
measurement_create_incrementality_analysis
ChatGPTIncrementality (geo lift testing): start an analysis that fits a model and measures the lift of a geo test. Use for questions like: - "Analyse the results of my test." - "Measure the lift for this test." Not for: - Checking the groups before the test -> get_incrementality_test_preview - Registering the test -> create_incrementality_test Before calling: test_id from list_incrementality_tests, with dates and region groups final. model is a free string; the API has no allow-list and the only value in backend fixtures is ridge. Confirm the model name and parameters with the user; do not assume synthetic_control. Returns: the created analysis with analysis_id. Model fitting is asynchronous; poll list_incrementality_analyses.
measurement_create_incrementality_test
ChatGPTIncrementality (geo lift testing): register a geo lift test from finalized test and control regions and dates. Use for questions like: - "Create the test with these regions." - "Register the test starting on 1 March." Not for: - Exploring which regions to use -> create_incrementality_test_design - Starting the statistical analysis of a test -> create_incrementality_analysis Before calling: metric_name from list_incrementality_metrics; region groups from get_incrementality_design_candidates and get_incrementality_candidate_power. Present test_group, control_group, dates and planned_spend to the user before confirming. Dates must be ordered training_start <= test_start <= test_end, and no region can be in both groups. test_type is a free string; the API has no enum, so copy the value from an existing test unless told otherwise. Returns: the created test with test_id.
measurement_create_incrementality_test_design
ChatGPTIncrementality (geo lift testing): start a design search that finds the best test and control region sets for a geo lift test. Use for questions like: - "Design a geo test for Facebook revenue." - "Find which regions to test in." - "Set up a test that can detect a 5% lift over 4 to 6 weeks." Not for: - Registering a test that is already decided -> create_incrementality_test - Checking whether the data is suitable first -> check_incrementality_data_readiness Before calling: metric_name from list_incrementality_metrics. Run check_incrementality_data_readiness first; do not continue on STOP. Effect sizes are relative lift from 0 to 0.4 (0.05 = 5%); the database rejects more than 0.4. Present duration range, effect bounds and region rules to the user before confirming. included_regions must appear in every candidate test group. excluded_regions are dropped from the analysis entirely. removed_regions are kept out of the test group but can still act as donors. Returns: the created design with design_id. The search is asynchronous; poll list_incrementality_test_designs (view="status") until SUCCEEDED, then call get_incrementality_design_candidates.
measurement_create_media_plan
ChatGPTMedia plans (budget optimiser and saved what-if scenarios): create a new media plan, either an optimised budget plan or a fixed what-if scenario that is saved and can be tracked. Use for questions like: - "Create a plan that maximises conversions with a 50k budget." - "Save this spend scenario as a plan." - "Set up a Q1 plan for Search and Paid Social." Not for: - Changing an existing plan -> update_media_plan - A quick unsaved what-if -> run_forecast_what_if Before calling: account_id and partition_id from list_accounts_and_partitions. Confirm the configuration with the user. Creating a plan never changes an existing one. Results appear only after run_media_plan. Config shape (flat keys): - Plan: name, scenario_type (optimized | fixed), active, start_date, end_date, reference_start_date, reference_end_date (YYYY-MM-DD) - Objective: optimisation_target (max_conversions | max_revenue | min_spend) - Restrictions at the top level, not nested: max_spend, min_spend, max_cpa, min_roas, min_conversions (null = unconstrained) - Channels: sources as a list of names (type auto, flight_pattern ref_period), OR source_budgets as full objects for per-channel type, min, max, flight_pattern. Not both. Example config: {"name": "Q1 test", "min_conversions": 500, "max_spend": 50000, "optimisation_target": "max_conversions", "sources": ["Search - Generic", "Paid Social"]} Returns: the created plan with its server-assigned id.
measurement_create_mmm_model_graph
ChatGPTMarketing Mix Modeling (MMM): create a new model graph from a name and a list of edges (media, control and target nodes). Use for questions like: - "Create a new MMM graph with Search and Social driving revenue." - "Copy this model structure into a new graph." Not for: - Changing an existing graph -> update_mmm_model_graph - Training a model on a graph -> start_mmm_experiment Before calling: get valid node names from list_mmm_available_graph_nodes and ask the user for the graph name. Existing names are visible in list_mmm_model_graphs, but duplicates are allowed. Send the full edge list in one call and show the name and edges to the user before confirming. A target node needs model_target, and an edge with adstock=true needs adstock_type. Adstock and lag belong on the edges, not on the experiment run. Effect: also creates a calibration row per media source, usually include=false, so get_mmm_graph_active_calibrations may return an empty list until those are enabled. Returns: the created graph with its id.
measurement_delete_incrementality_analysis
ChatGPTIncrementality (geo lift testing): soft-delete one analysis of a geo test. Use for questions like: - "Remove this failed analysis." - "Delete the old analysis and rerun it." Not for: - Deleting the whole test -> delete_incrementality_test Before calling: analysis_id from list_incrementality_analyses. Confirm which test it belongs to first. Effect: the analysis config is removed and its result endpoints may stop working. Returns: a deletion confirmation.
measurement_delete_incrementality_test
ChatGPTIncrementality (geo lift testing): soft-delete a geo lift test together with all its analyses. Use for questions like: - "Delete this test." - "Remove the test I set up by mistake." Not for: - Removing only one analysis -> delete_incrementality_analysis Before calling: test_id from list_incrementality_tests. The unconfirmed preview lists the analyses that will also be deleted; show that list to the user before confirming. Effect: the test, every analysis config on it and those analyses' runs get deleted_at set. Result files in S3 are kept. The test and analyses disappear from the list tools. Returns: a deletion confirmation.
measurement_deploy_mmm_trial
ChatGPTMarketing Mix Modeling (MMM): deploy a chosen trial as the live model for the partition. This replaces the current production model. Use for questions like: - "Deploy trial 00ebc736." - "Put this model into production." Not for: - Viewing deployments -> list_mmm_deployments, get_latest_mmm_deployment - Training -> start_mmm_experiment Before calling: trial_id from get_mmm_trial_metrics_table, after comparing trials and confirming the choice with the user. Check the current deployments first and say which trial will replace the active one and why. Optional body: full_refresh (default false) and truncation_days (default 45). Set full_refresh=true after changing between daily and weekly, or between source and campaign granularity; otherwise the API returns 400. Returns: the deployment with id, graph_id, experiment_id, trial_id and timestamps. It has no status field yet; use list_mmm_deployments to follow progress.
measurement_get_forecast_model_accuracy
ChatGPTForecasting (what-if on the deployed MMM model): accuracy of the model behind the forecasts, with fit metrics plus an actual-vs-predicted backtest over the training history. Use for questions like: - "How accurate is the forecast model?" - "Is the model biased or overfitting?" - "How well did the model track actuals month by month?" Not for: - Comparing candidate trials of a training run -> get_mmm_trial_metrics_table - Marketing ROAS, CPA or spend reporting -> get_performance_report Before calling: metric is the deployed graph's target: get_latest_mmm_deployment -> get_mmm_model_graph -> model_target. If that deployment is not SUCCEEDED, take the graph from the newest SUCCEEDED row in list_mmm_deployments. This is a best guess: the forecasting service loads the latest stored model, which may not be the latest successful deployment. Never assume "revenue". Set weekly=true for weekly models (holdout is 5 steps weekly, 15 daily). Returns: fit_metrics (r2, train_mape, holdout_mape, train_nrmse, holdout_nrmse) and predictions (n_points, date range, means, mean_abs_pct_error, bias_pct, monthly[]). This is a summary, not raw daily rows. Either section may be {"error": ...} if its backend call failed.
measurement_get_incrementality_actual_vs_counterfactual
ChatGPTIncrementality (geo lift testing): get the actual vs predicted (counterfactual) metric series for an analysis over its full history. Use for questions like: - "What would have happened without the campaign?" - "Plot actual against the counterfactual." Not for: - The lift numbers themselves -> get_incrementality_daily_lift - Per-day actual, predicted and effect together -> get_incrementality_analysis_daily_results Before calling: analysis_id from list_incrementality_analyses; the analysis must be finished. Returns: rows {date, kpi_name, kpi_value}, where kpi_name is Actual or Predicted.
measurement_get_incrementality_analysis_daily_results
ChatGPTIncrementality (geo lift testing): get the per-day results table of an analysis: actual, predicted, effect and cumulative effect. Use for questions like: - "What is the cumulative lift of the test?" - "Show me the daily results." Not for: - The lift with confidence bounds -> get_incrementality_daily_lift - Actual vs predicted series only -> get_incrementality_actual_vs_counterfactual Before calling: analysis_id from list_incrementality_analyses; the analysis must be finished. Returns: one row per day with actual, predicted, effect and cumulative_effect. Output is capped at 512KB. For cumulative lift only: sql="SELECT date, cumulative_effect FROM data".
measurement_get_incrementality_candidate_power
ChatGPTIncrementality (geo lift testing): show how likely one candidate test group is to detect a lift (power curve and minimum detectable effect). Use for questions like: - "Is this test powerful enough?" - "What is the smallest lift we could detect?" - "How long should the test run for 80% power?" Not for: - Ranking candidates -> get_incrementality_design_candidates - Results of a running or finished test -> the analysis tools Before calling: design_id with status SUCCEEDED. Optional test_group (comma-separated regions) selects a candidate; omit it for the top-ranked candidate. Returns: power_curve[], false_positive_rate, mde_at_80pct_power and power_by_duration_mde80, computed from the design simulations. Do not call on a design that is still RUNNING.
measurement_get_incrementality_daily_lift
ChatGPTIncrementality (geo lift testing): get the daily lift of a test with lower and upper bounds over the test window. Use for questions like: - "How much lift did the test create each day?" - "Was the effect significant?" Not for: - Actual vs counterfactual series -> get_incrementality_actual_vs_counterfactual - Cumulative lift -> get_incrementality_analysis_daily_results Before calling: analysis_id from list_incrementality_analyses; the analysis must be finished. Returns: rows with kpi_value and lower/upper bounds. kpi_value is the absolute effect in metric units, not a percentage. The confidence level of the bounds is not defined by the API. Output is capped at 512KB.
measurement_get_incrementality_design_candidates
ChatGPTIncrementality (geo lift testing): get the ranked candidate test/control region sets that a finished design produced. Use for questions like: - "Which regions should I test in?" - "What are the top candidates from my design?" Not for: - How likely a candidate is to detect a lift -> get_incrementality_candidate_power - Registering the chosen test -> create_incrementality_test Before calling: design_id from list_incrementality_test_designs or create_incrementality_test_design, with status SUCCEEDED (check list_incrementality_test_designs(view="status")). Returns: candidates ranked by final_rank, with test and control regions and scores, wrapped as {candidates, total_count, has_more}. Use limit and offset to page; sql= selects from the table data. Review the top candidates and get_incrementality_candidate_power before create_incrementality_test.
measurement_get_incrementality_geo_data
ChatGPTIncrementality (geo lift testing): fetch the raw geo-level input data (one row per date, region, metric) behind tests and designs. Use for questions like: - "Show me daily revenue for Texas." - "Which regions have zero days or gaps in the data?" - "Run a custom query over the geo data." Not for: - A go/no-go verdict on whether the data is good enough -> check_incrementality_data_readiness - Per-region share of volume -> get_incrementality_region_volume_share Before calling: metric_name from list_incrementality_metrics when filtering to one metric. Returns: rows {date, region, metric_name, metric_value}. The data is sparse: a date/region pair with no row has no data. The API has no row maximum; the tool fetches up to 1e6 rows and filters by metric_name itself. Output is capped at 512KB; use sql= to filter or aggregate. Example sql (table data, columns date, region, metric_name, metric_value): - One region: SELECT date, metric_value FROM data WHERE region = 'Texas' ORDER BY date - Daily cross-region mean: SELECT date, avg(metric_value) AS avg FROM data GROUP BY date ORDER BY date - Zero rate by region: SELECT region, avg(CASE WHEN metric_value = 0 THEN 1.0 ELSE 0 END) AS zero_rate FROM data GROUP BY region ORDER BY zero_rate DESC
measurement_get_incrementality_region_correlation
ChatGPTIncrementality (geo lift testing): measure how closely regions move together, to judge whether good control regions exist for a geo test. Use for questions like: - "Do our regions move together enough for a geo test?" - "Which regions are the best donors for a test region?" Not for: - A single go/caution/stop verdict -> check_incrementality_data_readiness - Raw region series -> get_incrementality_geo_data Before calling: metric_name from list_incrementality_metrics. Returns (sections default to summary, factors, donors; matrix is opt-in because it is large): - matrix: full pairwise correlation matrix - summary: mean and median comovement, histogram, badge - factors: PCA on the correlation matrix - donors: best and worst donor regions by average correlation
measurement_get_incrementality_region_volume_share
ChatGPTIncrementality (geo lift testing): show each region's share of your own total metric volume across the full history. Use for questions like: - "Which regions drive most of our revenue?" - "How concentrated is our volume across regions?" Not for: - Competitor market share (not available) - Whether the data suits a geo test -> check_incrementality_data_readiness Before calling: metric_name from list_incrementality_metrics. Returns: {metric_name, total, regions: [{region, value, share_pct}]}.
measurement_get_incrementality_test
ChatGPTIncrementality (geo lift testing): get the full configuration of one geo lift test (source, dates, test and control regions, planned spend). Use for questions like: - "What are the settings of this test?" - "Which regions are in test and control?" Not for: - Test results -> get_incrementality_actual_vs_counterfactual, get_incrementality_daily_lift - Raw series before the test -> get_incrementality_test_preview Before calling: test_id from list_incrementality_tests. Returns: the test configuration object.
measurement_get_incrementality_test_design
ChatGPTIncrementality (geo lift testing): get the full configuration of one test design (metric, duration range, effect-size range, region rules). Use for questions like: - "What settings did I use for this design?" - "Which regions did I exclude from the design?" Not for: - Job status of designs -> list_incrementality_test_designs (view="status") - Candidate region sets -> get_incrementality_design_candidates Before calling: design_id from list_incrementality_test_designs or create_incrementality_test_design. Returns: the design configuration object.
measurement_get_incrementality_test_preview
ChatGPTIncrementality (geo lift testing): preview the metric series for a test's pre-period and test window to check test and control groups fit well. Use for questions like: - "Do the test and control groups move together before the test?" - "Show the metric per region during the test." Not for: - Test results and lift -> get_incrementality_actual_vs_counterfactual, get_incrementality_daily_lift - Raw data for any region -> get_incrementality_geo_data Before calling: test_id from list_incrementality_tests. Returns: by=group: rows {date, group_name, metric_value} for test vs control aggregates. by=region: rows {date, region, metric_value} for every region, not only the test and control groups. All dates are returned, not just the test window. Output is capped at 512KB; use sql= to filter.
measurement_get_latest_mmm_deployment
ChatGPTMarketing Mix Modeling (MMM): get the most recent deployment attempt in this partition, which may have FAILED, with its graph and experiment ids. Use for questions like: - "What was deployed last?" - "Which graph is the deployed model built on?" Not for: - The model currently serving. The UI uses /deployments/successful-latest, a different endpoint, so check status here before assuming the latest row is live. - Deployment history -> list_mmm_deployments Before calling: account_id and partition_id from list_accounts_and_partitions. Chain get_latest_mmm_deployment -> get_mmm_model_graph -> model_target to find the metric for the forecasting tools. Returns: {graph_id, experiment_id, trial_id, status (null if unknown), graph_name, experiment_name, triggered_*}.
measurement_get_media_plan
ChatGPTMedia plans (budget optimiser and saved what-if scenarios): the full configuration of one media plan: dates, reference window, objective, restrictions and per-channel budgets. Use for questions like: - "What are the settings of this plan?" - "Which channels and limits does the Q1 plan use?" Not for: - The plan's results -> get_media_plan_projected_results, get_media_plan_pacing - Listing plans -> list_media_plans Before calling: media_plan_id from list_media_plans: a UUID, or the aliases default / default_f30. Call this before update_media_plan. Returns: the plan JSON with window_status, optimisation_target and reference_window_relationship, plus flat_config. Use flat_config (not the nested restrictions or source_budgets) as the shape for update_media_plan.
measurement_get_media_plan_channel_allocation
ChatGPTMedia plans (budget optimiser and saved what-if scenarios): how a plan shifts budget between channels: optimal vs reference vs actual spend per channel. Use for questions like: - "Where does the plan move budget?" - "How much is planned for Paid Social vs last period?" - "Which channels are under-spending against the plan?" Not for: - Plan totals -> get_media_plan_pacing - Channel names for reporting -> list_performance_channels_and_sources Before calling: media_plan_id from list_media_plans: a UUID, or the aliases default / default_f30. Returns: rows per source/channel with optimal_spend, reference_spend, delta, actual_spend and attainment (capped at 512KB). {"table": "not_available"} when the plan has no results.
measurement_get_media_plan_daily_timeline
ChatGPTMedia plans (budget optimiser and saved what-if scenarios): the day-by-day series of one plan for spend, revenue or conversion. Use for questions like: - "Show planned daily spend for this plan." - "How did revenue track against the plan day by day?" Not for: - Totals -> get_media_plan_pacing - Reported performance outside plans -> get_performance_report Before calling: media_plan_id from list_media_plans: a UUID, or the aliases default / default_f30. metric is spend, revenue or conversion (singular, as the API uses). Returns: metric=spend gives the planned (optimal) spend only, not realised spend. metric=revenue and metric=conversion give optimal, reference and actual series, with reference dates shifted onto the plan calendar. {"<metric>": "not_available"} when the plan has no results.
measurement_get_media_plan_pacing
ChatGPTMedia plans (budget optimiser and saved what-if scenarios): actual performance to date against the plan (pacing and attainment). Use for questions like: - "Is the plan on track?" - "How much of the planned revenue have we achieved so far?" - "Actual vs plan for this plan." Not for: - The projected outcome only -> get_media_plan_projected_results - Per-channel pacing -> get_media_plan_channel_allocation Before calling: media_plan_id from list_media_plans: a UUID, or the aliases default / default_f30. Returns: rows per metric with optimal (full plan window), actual (plan start to today, capped at the end date), reference (reference window) and attainment (actual / optimal to date). {"totals": "not_available"} when the plan has no results.
measurement_get_media_plan_projected_results
ChatGPTMedia plans (budget optimiser and saved what-if scenarios): the projected outcome of a plan's last run for the whole plan window: spend, conversions, revenue, CPA and ROAS. Use for questions like: - "What does the optimiser expect this plan to deliver?" - "What ROAS is the plan projected to reach?" Not for: - Actual vs plan to date (pacing) -> get_media_plan_pacing - Per-channel budgets -> get_media_plan_channel_allocation Before calling: media_plan_id from list_media_plans: a UUID, or the aliases default / default_f30. The plan must have run (see get_media_plan_run_status). Returns: projected metrics. {"metrics": "not_available"} when the plan has no results yet.
measurement_get_media_plan_run_status
ChatGPTMedia plans (budget optimiser and saved what-if scenarios): the status of a plan's latest run (optimiser or fixed what-if). Use for questions like: - "Has the plan finished running?" - "Why did the optimisation fail?" Not for: - The results of the run -> get_media_plan_projected_results, get_media_plan_pacing Before calling: media_plan_id from list_media_plans: a UUID, or the aliases default / default_f30. Poll this after run_media_plan. Returns: the run record with status, execution metadata and error_message when it failed, or {"status": "not_run"} when the plan has never run.
measurement_get_mmm_calibration_attribution_data
ChatGPTMarketing Mix Modeling (MMM): get conversions and revenue per source under each attribution method, to choose calibration targets. Use for questions like: - "What does platform attribution say about Search vs the MTA?" - "Show last-click conversions per source." Not for: - Which methods a source supports -> list_mmm_calibration_options - The calibration currently applied -> get_mmm_graph_active_calibrations Before calling: account_id and partition_id from list_accounts_and_partitions (this tool is partition-level; it takes no graph_id). granularity is daily or weekly. a2_source narrows to one source and must be an exact name as it appears in this tool's a2_source column: call once without it (e.g. sql="SELECT DISTINCT a2_source FROM data") to get the names, then filter. Returns: a capped table with ds, a2_source and platform, mta and lastclick conversions and revenue. Use sql= to filter.
measurement_get_mmm_experiment_config
ChatGPTMarketing Mix Modeling (MMM): get the training configuration an experiment was run with (model, dates, bounds and options). Use for questions like: - "What settings did this experiment use?" - "Which model and date range was it trained on?" Not for: - Trial results -> get_mmm_trial_metrics_table - Geo test designs -> get_incrementality_test_design Before calling: experiment_id from list_mmm_experiments. Returns: the experiment's configuration fields as submitted to start_mmm_experiment.
measurement_get_mmm_graph_active_calibrations
ChatGPTMarketing Mix Modeling (MMM): get the calibration constraints that are switched on (include=true) for one model graph. Use for questions like: - "Which sources are calibrated on this graph, and against which method?" - "What is the current calibration direction and type?" Not for: - The full list of calibratable sources -> list_mmm_calibration_options - The attribution data used for calibration -> get_mmm_calibration_attribution_data Before calling: graph_id from list_mmm_model_graphs. Calibrations are created with the graph and belong to it. Returns: list of {source, include, direction, metric, type, user_input_value}. The API hides include=false nodes, so this is not the full stored config, and an empty list is normal on a new graph. Sending this list back to replace_mmm_graph_calibration_config is safe: the hidden pairs are refilled as include=false defaults (any non-default metric or type they held is reset, as in the a2-frontend).
measurement_get_mmm_input_data_freshness
ChatGPTMarketing Mix Modeling (MMM): get when the MMM input data file was last updated. Use for questions like: - "How recent is the MMM data?" - "When was the model input last refreshed?" Not for: - Deployment or training job state -> list_mmm_deployments, list_mmm_experiments Before calling: account_id and partition_id from list_accounts_and_partitions. Returns: {last_modified}, only a timestamp. It says nothing about pipeline or job status.
measurement_get_mmm_input_data_statistics
ChatGPTMarketing Mix Modeling (MMM): get statistics about the data that feeds the MMM: overview, correlations, spend totals, per-source breakdown and activity windows. Use for questions like: - "Give me an overview of the MMM input data." - "What did we spend per source in Q1, as the model sees it?" - "Which inputs are correlated?" - "When was each source active?" Not for: - Raw feature series -> get_mmm_input_timeseries - Blended spend, revenue, ROAS and CPA reporting -> get_performance_report (blended=True) - How fresh the data is -> get_mmm_input_data_freshness Before calling: date_from and date_to are required for every section except summary; ask the user for the window if it is not given. get_mmm_input_data_freshness only returns a last-modified timestamp, not the data range. sections: summary, correlations, cost, breakdown, activity (omit for all). Prefer sections=["breakdown"] alone for per-source costs: cost.total_period_cost equals the sum of breakdown[].total_cost, so fetch both only if you also need the aggregate KPIs. Returns: one key per requested section. A section that fails comes back as {"error": ...} without failing the others.
measurement_get_mmm_input_timeseries
ChatGPTMarketing Mix Modeling (MMM): get the MMM input features (spend, controls, target) as a time series for a date window. Use for questions like: - "Show weekly spend for Search over 2025." - "How does the target trend against media spend?" - "Is there seasonality in the inputs?" Not for: - A trial's fitted vs actual line -> get_mmm_trial_fitted_vs_actual - Summary statistics -> get_mmm_input_data_statistics Before calling: date_from and date_to are required; ask the user for the window if it is not given. granularity is daily or weekly; normalize=true scales each source to 0..1. Returns: a capped table with columns ds, metric, source (one row per date and feature). Use sql= (SELECT on data).
measurement_get_mmm_model_graph
ChatGPTMarketing Mix Modeling (MMM): get one model graph with its edges, nodes, notes and AI summary. Use for questions like: - "What does this model graph look like?" - "Which channels feed the revenue target?" - "What is the model target (conversions or revenue)?" Not for: - Listing graphs or finding a graph by name -> list_mmm_model_graphs - The catalog of nodes you may use -> list_mmm_available_graph_nodes Before calling: graph_id from list_mmm_model_graphs (match by name first). Call this before update_mmm_model_graph, and to read model_target for the forecasting tools. Returns: the graph with nodes embedded on each edge (source, destination, sign, adstock, adstock_type, lag, saturation_curve), plus ai_summary and user_notes.
measurement_get_mmm_trial_channel_breakdown
ChatGPTMarketing Mix Modeling (MMM): get the per-channel breakdown of one trial: CPA, attribution, spend and their shares. Use for questions like: - "How did each channel perform in this trial?" - "What share of spend and attribution does Search have?" - "What is the CPA per channel for the last 90 days?" Not for: - Ranking trials -> get_mmm_trial_metrics_table - The fitted vs actual series -> get_mmm_trial_fitted_vs_actual Before calling: trial_id from get_mmm_trial_metrics_table (8 to 40 letters, digits, _ or -), plus experiment_id and graph_id. Returns: one row per triangulation_source with cpa, attributions, spend, share_of_spend and share_of_attribution, plus last_90_days_{channel} rows and a baseline row. The backend fills cpa (and mcpa) for conversion-target models and roas (and mroas) for revenue-target models, so the other one is null by design: use cpa for conversions, roas for revenue. adstock, lag and mu are often null.
measurement_get_mmm_trial_fitted_vs_actual
ChatGPTMarketing Mix Modeling (MMM): get the fitted vs actual series of one trial, in-sample and out-of-sample, to judge how well the model fits. Use for questions like: - "Show the model fit for this trial." - "How does the prediction compare to actuals in the holdout?" Not for: - Summary fit metrics across trials -> get_mmm_trial_metrics_table - The raw input series -> get_mmm_input_timeseries Before calling: trial_id from get_mmm_trial_metrics_table, plus experiment_id and graph_id. Optional date_from and date_to narrow the window. Returns: rows {date, type, value}, where type is one of in_sample_predictions, in_sample_actual, out_of_sample_predictions, out_of_sample_actual.
measurement_get_mmm_trial_metrics_table
ChatGPTMarketing Mix Modeling (MMM): get one row of fit metrics per trial of an experiment, to rank and compare trials and choose one to deploy. Use for questions like: - "Which trial is best?" - "Rank the trials by holdout MAPE." - "Compare CPA for Search across all trials." Not for: - One trial's per-channel CPA, spend and shares -> get_mmm_trial_channel_breakdown (use this table to compare or rank trials, or one metric across many trials) - One trial's fitted vs actual series -> get_mmm_trial_fitted_vs_actual Before calling: experiment_id from list_mmm_experiments; check that graph_id matches the experiment's graph. The table is wide (about 173 columns). Use sql= DESCRIBE data, or SELECT only the columns you need. - List trial ids: sql="SELECT trial_id FROM data". - One-trial scorecard: SELECT r2, holdout_mape, ... FROM data WHERE trial_id = '00ebc736'. Display labels: {_scorecard_labels} - Filter ranges: sql="SELECT min(r2), max(r2), min(holdout_mape), max(holdout_mape) FROM data". - Drop exploded trials (r2 hugely negative or MAPE in the millions) before ranking. Returns: a capped table. Core columns: trial_id, AI_selection (1 for trials the best-trials service recommends; all 0 if that service is unavailable), r2, train_mape, holdout_mape, holdout_mape_30/60/90_days, nrmse_, mae_, decomp_rssd, baseline_share_attributed, zero_weights_regressors, performance_intersection/left/missing_source_count, calibration_, config/conversions_, coef__beta, baseline_weekly_sin_. Per channel: CPA_{{channel}}, {{channel}}_attributions, {{channel}}_roi (not ROAS; always populated, so use it to compare a channel across trials). Quote hyphenated names, e.g. "CPA_Search_-_Branded". trial_id values are 8-character hex strings (e.g. 00ebc736).
measurement_get_performance_report
ChatGPTPerformance reporting (marketing results): spend, revenue, conversions, ROAS and CPA for a date range, broken down by channel, source or campaign, or as blended all-channel totals. Use for questions like: - "How did Paid Social perform last month?" - "What was our total ROAS in Q2?" - "Show weekly spend and revenue by channel." - "Which campaigns in Search - Generic had the best CPA?" Not for: - What-if or future spend -> run_forecast_what_if - Diminishing returns for one source -> get_source_saturation_curve - MMM model statistics -> get_mmm_input_data_statistics Before calling: exact channel and source names from list_performance_channels_and_sources. Drill-down is channel -> source -> campaign: sources requires channels and campaigns requires sources. aggregation_level: full_period (default), daily, weekly, monthly, yearly. - blended=False (default): a row table with row-level ratios. sql= runs DuckDB on table data (~33 columns; sql="DESCRIBE data" lists them). Spend is spend, not a2_cost. Metric families use the prefixes a2_, mta_, platform_, firstclick_, lastclick_. Baseline appears as a row labelled "Baseline" with no spend. - blended=True: all channels summed; media-only totals plus baseline_ fields and both media-only and *_including_baseline ratios. Cannot be combined with channels, sources, campaigns or sql. Returns: blended=False: a table capped at 512KB. blended=True: one totals dict (full_period) or a list of {date, ...} for other levels.
measurement_get_source_saturation_curve
ChatGPTPerformance reporting (marketing results): the saturation (response) curve for one media source, showing diminishing returns as spend increases. Use for questions like: - "Where does Search - Generic hit diminishing returns?" - "What would an extra 10k on Paid Social bring in?" - "Show the response curve for Display Brand." Not for: - Actual results for a period -> get_performance_report - A full budget scenario across channels -> run_forecast_what_if Before calling: the exact source name from list_performance_channels_and_sources. Returns: a list of spend levels with incremental conversions, CPA, revenue and ROAS for the source and date window.
measurement_list_accounts_and_partitions
ChatGPTMeasurement scope (accounts and partitions): list the Funnel Measurement accounts the caller can use, each with its partitions. Call this first to get account_id and partition_id for every other measurement tool. These are Measurement accounts only: the subset of the caller's accounts that have Measurement set up and are active. Other Funnel accounts, and ad-platform accounts such as Google Ads or Meta, are not listed and cannot be used as account_id. Use for questions like: - "Which accounts and markets can I analyse?" - "Use our German revenue partition." - "Which brands do we measure?" Not for: - Model graphs or experiments -> list_mmm_model_graphs, list_mmm_experiments - Channels or sources -> list_performance_channels_and_sources Before calling: nothing. By default only active partitions are returned (active=true); active=None returns all. A partition is one brand x market x conversion event, with its own currency and time zone. Match the user's description against brand, market, currency, conversion and time_zone before asking them to pick an id. Returns: list of {id, name, active, partitions: [{id, name, active, brand, market, currency, conversion, time_zone}]}. Pass the account id as account_id and the partition id as partition_id. Text fields are omitted when the config API does not send them.
measurement_list_forecast_channels
ChatGPTForecasting (what-if on the deployed MMM model): the spend channels and on/off flags that can be changed in a forecast scenario, with their spend. Use for questions like: - "Which channels can I change in a what-if?" - "What will we spend on Paid Social next month, per the model?" - "What date range does the forecast model cover?" Not for: - Channels for performance reporting -> list_performance_channels_and_sources - Channels in an MMM graph -> list_mmm_available_graph_nodes - Channel budgets in a media plan -> get_media_plan_channel_allocation Before calling: metric is the deployed graph's target: get_latest_mmm_deployment -> get_mmm_model_graph -> model_target. If that deployment is not SUCCEEDED, take the graph from the newest SUCCEEDED row in list_mmm_deployments. This is a best guess: the forecasting service loads the latest stored model, which may not be the latest successful deployment. Never assume "revenue". start_date and end_date are optional, must be set together, and scope the spend totals (end at most 730 days from today). Use them to read the exact window spend before expressing a change as a percentage. Returns: channels[] with spend totals, binary_features[], min_date, max_date. Future windows show projected spend based on last year's pattern. Use these exact names in run_forecast_what_if: unknown channel or flag names are silently ignored by the service.
measurement_list_incrementality_analyses
ChatGPTIncrementality (geo lift testing): list the analyses (model fits that measure lift) run on tests in this partition. Use for questions like: - "Has the analysis for my test finished?" - "Which analyses exist for this test?" Not for: - The tests themselves -> list_incrementality_tests - The lift results -> get_incrementality_actual_vs_counterfactual, get_incrementality_analysis_daily_results, get_incrementality_daily_lift Before calling: account_id and partition_id from list_accounts_and_partitions. Optional analysis_id returns only that analysis. Returns: list of analyses with analysis_id, test, model and status.
measurement_list_incrementality_metrics
ChatGPTIncrementality (geo lift testing): list the metrics (e.g. revenue, conversions) that geo tests can be run on for this partition. Use for questions like: - "Which metrics can I run an incrementality test on?" - "What KPIs are available for geo testing?" Not for: - Metric values or time series -> get_incrementality_geo_data - Metrics for MMM modelling -> the modeler tools Before calling: account_id and partition_id from list_accounts_and_partitions. Call this first: every other incrementality tool that takes metric_name needs a value from here. Do not assume "revenue" exists. Returns: list of metric name strings.
measurement_list_incrementality_regions
ChatGPTIncrementality (geo lift testing): list the regions (geo units, e.g. states or DMAs) that have data for a metric. Use for questions like: - "Which regions do we have data for?" - "What regions can I put in a test or control group?" Not for: - Ranked test/control region sets -> get_incrementality_design_candidates - Per-region volume -> get_incrementality_region_volume_share Before calling: metric_name from list_incrementality_metrics. Returns: list of region names.
measurement_list_incrementality_test_designs
ChatGPTIncrementality (geo lift testing): list the test designs (searches for the best test/control region split) in this partition. Use for questions like: - "What test designs have I created?" - "Has my design finished running?" - "Which designs succeeded?" Not for: - Running or planned tests -> list_incrementality_tests - The ranked region sets a design produced -> get_incrementality_design_candidates Before calling: account_id and partition_id from list_accounts_and_partitions. Returns: view=status (default): batch job status plus metadata per design, as in the a2-frontend table. Wait for SUCCEEDED before get_incrementality_design_candidates. view=full: full design configuration without batch status; use get_incrementality_test_design for one design.
measurement_list_incrementality_tests
ChatGPTIncrementality (geo lift testing): list the geo lift tests registered in this partition (planned, running or finished). Use for questions like: - "What incrementality tests do we have?" - "Which tests ran on Facebook?" - "Show me my geo tests." Not for: - Design searches -> list_incrementality_test_designs - Results of a test -> list_incrementality_analyses Before calling: account_id and partition_id from list_accounts_and_partitions. Returns: list of tests with id, name, type, source, dates and regions.
measurement_list_media_plans
ChatGPTMedia plans (budget optimiser and saved what-if scenarios): list the media plans in this partition. Use for questions like: - "Which media plans do we have?" - "Is there an active plan for Q3?" Not for: - Ranking plans by outcome -> compare_media_plans - One plan's full configuration -> get_media_plan Before calling: account_id and partition_id from list_accounts_and_partitions. detail="summary" (default) gives compact rows; detail="full" gives the raw planner collection (budgets, controls). Returns: summary rows {id, name, scenario_type, active, dates, window_status, optimisation_target, restrictions}, or the raw API JSON for detail="full".
measurement_list_mmm_available_graph_nodes
ChatGPTMarketing Mix Modeling (MMM): list every media, control and target node that can be used to build a model graph in this partition. Use for questions like: - "Which channels can I put in the model?" - "What control variables are available?" - "What target can I model?" Not for: - The nodes already in a graph -> get_mmm_model_graph - Spend in a date window -> get_mmm_input_data_statistics Before calling: account_id and partition_id from list_accounts_and_partitions. Use the names exactly as returned in create_mmm_model_graph and update_mmm_model_graph. Returns: list of {name, type (media, control, target), input_metric, model_target, position}. input_metric lists the available metrics for the node; only the synthetic "Model Target" node has model_target, also as a list.
measurement_list_mmm_calibration_options
ChatGPTMarketing Mix Modeling (MMM): list the sources on one model graph that can be calibrated, with the calibration methods each source supports. Use for questions like: - "Which sources can I calibrate?" - "Is incrementality available as a calibration method for Search?" Not for: - What is currently applied to a graph -> get_mmm_graph_active_calibrations - The attribution numbers behind each method -> get_mmm_calibration_attribution_data Before calling: graph_id from list_mmm_model_graphs. The catalog is graph-scoped: only media sources on that graph, the same list the a2-frontend calibration modal uses. Returns: list of {source, metrics}, where metrics are the accepted methods (platform, mta, lastclick, incrementality, user_input, external; user_input is always present). These are the only sources replace_mmm_graph_calibration_config accepts.
measurement_list_mmm_deployments
ChatGPTMarketing Mix Modeling (MMM): list all deployment jobs in this partition with their status. Use for questions like: - "What has been deployed and when?" - "Did the last deployment succeed?" Not for: - The latest row with graph_id and experiment_id -> get_latest_mmm_deployment Before calling: account_id and partition_id from list_accounts_and_partitions. Check this before deploy_mmm_trial. Returns: rows of {id, graph_name, experiment_name, trial_id, status, triggered_at, ended_at}.
measurement_list_mmm_experiments
ChatGPTMarketing Mix Modeling (MMM): list MMM experiments (model-training jobs) in this partition, optionally for one graph. Use for questions like: - "Has my training run finished?" - "Which experiments succeeded or failed?" - "What runs exist for this graph?" Not for: - Geo lift tests -> list_incrementality_tests - The trials of one experiment -> get_mmm_trial_metrics_table Before calling: account_id and partition_id from list_accounts_and_partitions. Optional graph_id limits to one graph. Each row carries graph_id and graph_name, so it can also resolve an experiment_id from a graph name. Returns: rows of {id, graph_id, graph_name, job_name, model_name, iterations, weekly, job_started_by, job_status, created_at, started_at, ended_at}.
measurement_list_mmm_model_graphs
ChatGPTMarketing Mix Modeling (MMM): list the model graphs (the saved model specifications: which channels drive which target) in this partition. Use for questions like: - "Which MMM models do we have?" - "Find the graph called 'Q3 brand model'." - "How many experiments and deployments does each graph have?" Not for: - Partitions or workspaces -> list_accounts_and_partitions - The structure of one graph -> get_mmm_model_graph - Training runs -> list_mmm_experiments Before calling: account_id and partition_id from list_accounts_and_partitions. This is the way to turn a graph name into a graph_id; a name from the user beats a stale graph_id from earlier in the conversation. Returns: list of {id, name, experiments_count, deployments_count}.
measurement_list_performance_channels_and_sources
ChatGPTPerformance reporting (marketing results): the valid channel and source names for this partition, and which sources belong to which channel. Use for questions like: - "Which channels do we have?" - "What sources are under Paid Social?" - "Is Search - Branded a channel or a source?" Not for: - Channels you can change in a forecast -> list_forecast_channels - Nodes for an MMM graph -> list_mmm_available_graph_nodes Before calling: account_id and partition_id from list_accounts_and_partitions. Call this first, before any other performance tool on the partition: it is the authoritative {channel, source} map. Do not guess channel vs source from general knowledge. Returns: list of {channel, source}.
measurement_replace_mmm_graph_calibration_config
ChatGPTMarketing Mix Modeling (MMM): set calibration rows for a model graph. Send only the rows you want active or changed; the tool completes and writes the full list (the underlying call is a full PUT). Calibration is set per source AND direction (upper, lower, equal). Use for questions like: - "Calibrate Search against incrementality results." - "Turn off calibration for Affiliate." - "Set a user-provided ROAS for Social." Not for: - Reading the current state -> get_mmm_graph_active_calibrations - Listing options -> list_mmm_calibration_options Before calling: graph_id from list_mmm_model_graphs, sources from list_mmm_calibration_options (graph-scoped), and the current active rows from get_mmm_graph_active_calibrations. Send only the rows you want to set. Like the a2-frontend, the tool builds the full list of one row per (catalog source, direction) and fills every pair you did not send with include=false defaults. The call is refused if a row that is active today is missing (send it to keep it, or with include=false to switch it off), if a source is not on the graph's catalog, if an included row uses a metric the source does not offer in list_mmm_calibration_options (user_input is always allowed), or if a (source, direction) pair is sent twice. It is also refused, with nothing written, if the catalog or current config cannot be read. The preview shows the final list; show it to the user before confirming. There is no weight field. Each node is source, include, direction (upper|lower|equal), metric (platform|mta|lastclick|incrementality|user_input|external), type (contribution|cpa|roas) and user_input_value (required when metric=user_input and include=true). Returns: the stored nodes as confirmed by the API.
measurement_run_forecast_what_if
ChatGPTForecasting (what-if on the deployed MMM model): forecast the metric for a date window, optionally under a what-if spend scenario, and return the impact. Nothing is saved. Use for questions like: - "What if we doubled Search spend next quarter?" - "What will revenue be next month?" - "What happens if we switch the promo flag off?" Not for: - A what-if you want to save, compare or track as a plan -> create_media_plan with scenario_type=fixed - Optimised budget allocation -> create_media_plan with scenario_type=optimized - Reported campaign ROAS or CPA -> get_performance_report Before calling: metric is the deployed graph's target: get_latest_mmm_deployment -> get_mmm_model_graph -> model_target. If that deployment is not SUCCEEDED, take the graph from the newest SUCCEEDED row in list_mmm_deployments. This is a best guess: the forecasting service loads the latest stored model, which may not be the latest successful deployment. Never assume "revenue". Call list_forecast_channels(metric, start_date, end_date) first for exact channel and flag names and the window spend. - modifications: absolute spend totals for the window, not multipliers. Unknown channel names are silently ignored. - binary_feature_overrides: on/off flags (all_on, all_off, flip), never in modifications. Unknown flag names are silently ignored. - flight_patterns (ref_period, previous_year, linear): how a modified channel's spend is spread over the window. Every key must also be in modifications. end_date is at most 730 days from today. Without modifications or overrides the tool returns the plain baseline forecast. Returns: baseline_total, scenario_total, actual_total (historical days in the window only), absolute_impact, pct_impact and window {start, end, days}.
measurement_run_media_plan
ChatGPTMedia plans (budget optimiser and saved what-if scenarios): run a media plan to compute its results. Use for questions like: - "Run the optimiser on this plan." - "Compute the results for my what-if plan." Not for: - Checking progress -> get_media_plan_run_status - An unsaved what-if -> run_forecast_what_if Before calling: media_plan_id from list_media_plans: a UUID, or the aliases default / default_f30. Confirm with the user; the preview shows which kind of run will start. - scenario_type=optimized starts the budget optimiser. - scenario_type=fixed runs a what-if scenario on the forecasting model, not the optimiser. If a run is already in progress, that run is returned and no new one starts. Runs cost backend compute and cannot be cancelled from here. Returns: the run record. Poll get_media_plan_run_status until it finishes.
measurement_start_mmm_experiment
ChatGPTMarketing Mix Modeling (MMM): start an experiment, meaning an asynchronous model-training job, on a model graph. Use for questions like: - "Train the model on this graph." - "Run an MMM with the ridge model from January." - "Start a new experiment with 2000 iterations." Not for: - Geo lift tests -> the incrementality tools - Checking on existing runs -> list_mmm_experiments Before calling: graph_id from list_mmm_model_graphs. Confirm the graph, model and training window with the user. Adstock and lag are set on the graph edges, not here. model_name defaults to orbit; set only the options relevant to the chosen model. saturation_type must be logistic or log for orbit, ridge and dlt, and logistic, michaelis_menten or hill for arbol. Omit optional fields to use the API defaults. calibrate=true applies the graph's calibration config. Returns: the experiment id and job metadata. Training is asynchronous; poll list_mmm_experiments until SUCCEEDED.
measurement_update_media_plan
ChatGPTMedia plans (budget optimiser and saved what-if scenarios): change an existing media plan's dates, objective, limits or channels. Use for questions like: - "Raise the spend cap to 60k." - "Remove the CPA limit." - "Add Display to the plan." Not for: - A new plan -> create_media_plan - Recomputing results -> run_media_plan (updates do not rerun the plan) Before calling: media_plan_id from list_media_plans: a UUID, or the aliases default / default_f30; read the current values with get_media_plan and show the exact change before confirming. Send only fields that change; omitted fields keep their stored values. Start from flat_config. - Restrictions at the top level; null on a restriction removes that limit. - Explicit null is refused on scenario_type, active, dates, optimisation_target and sources. - sources: channel names; existing channels keep their type, min and max when names match. - source_budgets: the full object list when changing per-channel type, min, max or flight_pattern. Example configs: {"min_conversions": 500} {"sources": ["Search - Generic", "Paid Social"]} {"max_spend": null} (remove the spend cap) Returns: the changed fields and the updated plan.
measurement_update_mmm_model_graph
ChatGPTMarketing Mix Modeling (MMM): rename a model graph, change its notes, or replace its edges. Use for questions like: - "Rename this graph." - "Add Paid Social to the model." - "Change the adstock on the Search edge." Not for: - A new graph -> create_mmm_model_graph - Changing training settings -> start_mmm_experiment Before calling: graph_id from list_mmm_model_graphs, and read the current state with get_mmm_model_graph. Omitted fields stay unchanged. Sending edges REPLACES every edge; it is not a merge, so send the complete list. The structure cannot be changed while this graph is the partition's latest successful deployment (the API returns 400); name and notes still can. Show the name and edge diff to the user before confirming. Returns: the updated graph.
mutate_fields
ChatGPTPrepare a create or update of 1 to 5 custom fields as one homogeneous proposal, without writing anything. Each field is checked for definition errors, the group is previewed once with sample values, and the result is a plan_id with the proposal (fields, before/after, dependency order, sample_data). Present the proposal and let the user pick which fields to apply; apply_field_mutation then writes the approved selection. Siblings reference each other as $key. To change some rules of an existing rule-based field, send rule_edits instead of the whole derived. Editing permission on fields is required, even for preparing. If you need rule or formula guidance, call list_guides with guide_id=field-rules before proposing changes; omit guide_id to list available guides.
prepare_add_dashboards_to_portal
ChatGPTProposes adding dashboards (ids from list_dashboards) to a portal. Each dashboard is read from dashboards-service in the workspace: dashboards are those that will be added, under their titles; already_in_portal and not_found will not be. Returns unchanged when none can be added. Changes nothing: returns change to show the user and a proposal for execute_portal_change. not_found when the workspace has no such portal. Needs get_account + edit_portals. Show the user change and ask them to approve it; only then call execute_portal_change with the proposal_id. Nothing is changed until then.
prepare_archive_budget
ChatGPTProposes archiving an active budget. Archiving deletes its budget data source, so its series leaves get_data and dashboards until it is restored with prepare_restore_budget. Changes nothing: returns change to show the user and a proposal for execute_budget_change. not_found when the workspace has no such active budget. Needs get_account, edit_plan_budgets, edit_fields and edit_datasources. Show the user change and ask them to approve it; only then call execute_budget_change with the proposal_id. Nothing is changed until then.
prepare_archive_convention
ChatGPTPropose archiving an active convention. Plan also archives its linked fields, campaigns and adherence. Show generated_dimensions (before/after) and the other effects to the user before approval. Nothing is written until execute_convention_change. Read list_guides(guide_id='conventions') first. Show the user the full change, including generated_dimensions before/after and linked-field effects; ask for approval. Only then call execute_convention_change with proposal_id. Nothing is changed until then.
prepare_create_brand
ChatGPTProposes creating a portal brand, a name and colours without a logo, for the subscription that owns the workspace. Changes nothing: returns change to show the user and a proposal for execute_brand_change. Needs get_account on the workspace and edit_subscription_portals on its subscription. Show the user change and ask them to approve it; only then call execute_brand_change with the proposal_id. Nothing is changed until then.
prepare_create_budget
ChatGPTProposes creating a Plan budget with a spend target and optional extra targets (up to 5), with no breakdown, as Plan's wizard would. The periods follow the cadence: calendar weeks (from the workspace's first day of the week), months or quarters, the first and last clipped to the budget dates; duration is one period for the whole budget; at most 53 periods. Give spend and each target's amounts either as one amount used in every period or as a list with one amount per period; a total to split across periods is not accepted. A target's field_id is a metric from the workspace's fields (search_fields), never a dimension; a formula field (for example CPC) is a fixed target: its amounts are per period, never summed into a total. Breakdowns, rollover, pacing, daily and custom cadences are set up in Plan. Changes nothing: returns change to show the user (the periods, and each target's amounts and total) and a proposal for execute_budget_change. Needs get_account, edit_plan_budgets, edit_fields and edit_datasources. Show the user change and ask them to approve it; only then call execute_budget_change with the proposal_id. Nothing is changed until then.
prepare_create_convention
ChatGPTPropose a Plan convention that may generate many dimensions from one input field. Each sourceType belongs to at most one pattern; one pattern may have omitted/empty sourceTypes as the all-other-connectors fallback. Review the complete definition and generated_dimensions before approval. Plan API creates the fields; nothing is written until execute_convention_change. Requires get_account, edit_plan_conventions and edit_fields. Read list_guides(guide_id='conventions') first. Show the user the full change, including generated_dimensions before/after and linked-field effects; ask for approval. Only then call execute_convention_change with proposal_id. Nothing is changed until then.
prepare_create_dashboard
ChatGPTPropose creating a new dashboard in the workspace titled title. To copy an existing dashboard, such as one market's or channel's dashboard for another, give its source_dashboard_id: without config the copy has the source's cards, layout, filters and date range as they are; with config it has that config instead (read the source with get_dashboard, change only what differs, such as its global filters, and pass the whole config). Without a source, config is the whole new dashboard: data tables, cards, layout, filters and date range, shaped as get_dashboard's config. A text card is {id, type: "text", content, position, title?}: content is the text to show, in Markdown. Changes nothing: checks the config and returns change, whose text_cards are the new dashboard's text cards with their content; when asking for approval, also say what else differs from the source, such as its filters. An invalid config, or a text card without content, is invalid_input. Show the user change and ask them to approve it; only then call execute_dashboard_change with the proposal_id. Nothing is changed until then.
prepare_create_dashboard_share_link
ChatGPTProposes a new public share link for a dashboard, as the Dashboards app's sharing dialog creates one: anyone who has the link can view the dashboard without signing in, until it is invalidated (prepare_invalidate_dashboard_share_link). Changes nothing: returns change to show the user and a proposal for execute_dashboard_share_link_change, which answers the new link's url. not_found when the workspace has no such dashboard. Needs get_account and edit_dashboards on the workspace. Show the user change: the dashboard that anyone with the new link can view without signing in. Ask them to approve it; only then call execute_dashboard_share_link_change with the proposal_id. No link exists until then.
prepare_create_data_source
ChatGPTPropose creating one Funnel data source from a completed setup (connection, remote account and template selected). Changes nothing: the setup is checked again and the exact create is signed. Show the user change (connector, connection, remote account, template and name) and ask them to approve it; only then call execute_create_data_source with the proposal_id. use_connect means this setup cannot be created here; the user sets it up in Connect.
prepare_create_export
ChatGPTProposes a new enabled, scheduled export of workspace data to a destination, and by default runs it once right after it is created. Changes nothing: returns change to show the user and a proposal for execute_export_change. destination.type picks the destination: gsheets (a Google Sheets tab, replaced on every run), bigquery (a table in the user's own Google Cloud project) or snowflake (a table in the user's own Snowflake account, in an existing database) or gcs (gzip CSV files in the user's own Cloud Storage bucket, one per month). A BigQuery or Snowflake table, or a Cloud Storage export's files, is written in monthly parts on the date dimension, which it must include, and kept up to date for the export's date range. Fields are checked against the workspace catalog: an unknown id, or one under the wrong role, is invalid_input; so is a connection that is not one of the user's (list_export_connections). Runs daily at schedule.hour:schedule.minute in schedule.time_zone, exactly as given. Needs get_account + edit_shares. Show the user change and ask them to approve it; only then call execute_export_change with the proposal_id. Nothing is changed until then.
prepare_create_portal
ChatGPTProposes creating a client portal in the workspace, without dashboards; add them with prepare_add_dashboards_to_portal once it exists. A brand_id must be one of the subscription's brands (list_brands). Changes nothing: returns change to show the user and a proposal for execute_portal_change. Needs get_account + edit_portals. Show the user change and ask them to approve it; only then call execute_portal_change with the proposal_id. Nothing is changed until then.
prepare_create_scheduled_task
ChatGPTProposes a scheduled task: a prompt that re-runs automatically on a schedule, as you, on this workspace. The result of each run is saved to a chat you can open from the run history. Use when the user asks to run something regularly ('every Monday', 'daily at 9', 'each morning'). schedule_expression: Use rate(N minutes|hours|days) e.g. rate(1 day), or a 6-field AWS cron cron(minute hour day-of-month month day-of-week year) with '?' for an unspecified day, e.g. cron(0 9 ? MON ) for every Monday 09:00. timezone is an IANA name (e.g. Europe/Stockholm); defaults to UTC. delivery_channels selects where each result is sent; omitting it defaults to email for this user. Pass [] for no notification. Use an explicit selection only when the user asks. Results go to the workspace members picked as email recipients, by default the task owner. Scheduled tasks can also be sent to other workspace members, but you cannot add, remove, or view those recipients: tell the user to manage them in the Scheduled Tasks page. Updating a task here keeps its existing recipients. Never say a task can only be sent to one person. slack_channel posts Slack results into a channel instead of the owner's DM. Pass the channel name the user gave (e.g. #marketing); it is looked up among the channels the Funnel AI app is in. Never ask for or quote a channel ID: refer to channels by name only. If the channel cannot be used, relay the returned message: the user may need to invite the app with /invite @Funnel AI in that channel, or confirm the name. If it is unclear which channel the user means (e.g. #marketing or #marketing-eu), ask the user to choose before setting it. Only the task owner can change it. When a task delivers to Slack and the user has not named a channel, mention once that results can go to a Slack channel instead of their DM. Confirm the name, prompt + cadence with the user first if ambiguous. Changes nothing: checks the task and returns change to show the user and a proposal for execute_scheduled_task_change. Needs get_account and edit_account. Show the user change and ask them to approve it; only then call execute_scheduled_task_change with the proposal_id. Nothing is changed until then.
prepare_create_view
ChatGPTProposes a new view of workspace data for a BI tool that reads from Funnel: datastudio (Looker Studio). Funnel writes nothing anywhere and needs no connection: the user connects the view in Looker Studio afterwards, with Funnel's connector, and picks dates there. Changes nothing: returns change to show the user and a proposal for execute_export_change. Fields are checked against the workspace catalog (an unknown id, or one under the wrong role, is invalid_input), and a formula metric needs the fields it is computed from in the view too. Needs get_account + edit_shares. Show the user change and ask them to approve it; only then call execute_export_change with the proposal_id. Nothing is changed until then.
prepare_create_workspace
ChatGPTProposes creating a workspace in the subscription that owns this workspace, with the user as its editor. Changes nothing: returns change to show the user and a proposal for execute_create_workspace. not_allowed says why the user cannot create one here. Needs get_account on the workspace and get_subscription and create_account on its subscription. Show the user change, including the subscription, and ask them to approve it; only then call execute_create_workspace with the proposal_id. Nothing is created until then.
prepare_data
ChatGPTLoads raw, row-level data without aggregation for custom analysis and re-slicing. scope controls the Data Source boundary: inferred resolves from requested fields, workspace accepts cross-platform fields, and data_source enforces one immutable ID. Omitted scope preserves pre-typed compatibility behavior during migration. Deprecated source_id remains accepted but cannot be combined with scope. When provided, the validated scope is applied to the MELD fetch, persisted with the table, and returned in metadata. Field discovery is platform-catalog evidence, not proof that every field is enabled by one exact source configuration.
prepare_delete_brand
ChatGPTProposes deleting a portal brand of the subscription that owns the workspace. Portals using it, in any workspace of the subscription, revert to the default branding; portals_using counts them now. There is no route to restore a brand. Changes nothing: returns change to show the user and a proposal for execute_brand_change. not_found when the subscription has no such brand. Needs get_account on the workspace and edit_subscription_portals on its subscription. Show the user change and ask them to approve it; only then call execute_brand_change with the proposal_id. Nothing is changed until then.
prepare_delete_field
ChatGPTPropose deleting one custom field, without deleting anything. Funnel archives the field: it disappears for everyone in the workspace, and only Funnel support can bring it back. Only custom fields can be deleted; built-in and connector fields answer not_found. Pass the field's exact id from search_fields; if the user's name for it matches more than one field, ask which one first. This tool cannot check where the field is used: the result says usage_checked: false. Show the user change and warning, and ask them to approve deleting it; only then call execute_field_change with the proposal_id.
prepare_delete_portal
ChatGPTProposes deleting a portal: its users lose access and its dashboards are untouched. There is no route to restore a portal, so treat it as permanent. Changes nothing: returns change to show the user and a proposal for execute_portal_change. not_found when the workspace has no such portal. Needs get_account + edit_portals. Show the user change and ask them to approve it; only then call execute_portal_change with the proposal_id. Nothing is changed until then.
prepare_delete_scheduled_task
ChatGPTProposes deleting (stopping) a scheduled task by id so it no longer runs. Get the id from list_scheduled_tasks. To only PAUSE it, use prepare_update_scheduled_task with status='disabled' instead. Changes nothing: returns change to show the user and a proposal for execute_scheduled_task_change; do not ask for a separate prose confirmation first. not_found when the workspace has no such task. Needs get_account and edit_account. Show the user change and ask them to approve it; only then call execute_scheduled_task_change with the proposal_id. Nothing is changed until then.
prepare_edit_budget
ChatGPTProposes changing an active budget's spend, its other targets or its dates, in one change. Read it with get_budget first: it shows the periods, each target's field_id and amounts, and whether it is editable. Give at least one of: spend (one amount used in every period, or one per period of the edited budget); set_targets (targets to add or change: a metric field_id from search_fields and amounts in the same two forms; a formula field such as CPC takes a value per period too, never summed); remove_targets (field ids; spend cannot be removed); start_date/end_date (the periods are rebuilt by the budget's cadence; a period with the same start date keeps its amounts unless new ones are given; when periods are added, give every remaining target, spend included, amounts in this call). Amounts are stored as given: a percent is a fraction (0.05 for 5%), a duration is in seconds. Renaming is prepare_update_budget. Budgets with a breakdown or rollover, budgets Plan's newer format cannot read, and date changes on daily or custom budgets are edited in Plan: invalid_input says why. Changes nothing: reads the budget and returns unchanged when nothing would change, otherwise change to show the user (every period and target, before and after) and a proposal for execute_budget_change. not_found when the workspace has no such active budget. Needs get_account, edit_plan_budgets, edit_fields and edit_datasources. Show the user change and ask them to approve it; only then call execute_budget_change with the proposal_id. Nothing is changed until then.
prepare_invalidate_dashboard_share_link
ChatGPTProposes invalidating one of a dashboard's active public share links (from list_dashboards' share_links), as the Dashboards app's sharing dialog does: anyone opening it afterwards is refused, and it cannot be brought back (create a new link in the app instead). Changes nothing: returns change to show the user and a proposal for execute_dashboard_share_link_change. not_found when the workspace has no such dashboard, or the dashboard has no such active link. Needs get_account and edit_dashboards on the workspace. Show the user change: the dashboard and the link that stops working for everyone who has it. Ask them to approve it; only then call execute_dashboard_share_link_change with the proposal_id. Nothing is invalidated until then.
prepare_remove_dashboards_from_portal
ChatGPTProposes removing dashboards from a portal; the dashboards themselves are untouched. dashboards are those the portal has and will lose, not_in_portal ids it does not have. Returns unchanged when the portal has none of them. Changes nothing: returns change to show the user and a proposal for execute_portal_change. not_found when the workspace has no such portal. Needs get_account + edit_portals. Show the user change and ask them to approve it; only then call execute_portal_change with the proposal_id. Nothing is changed until then.
prepare_remove_workspace_users
ChatGPTProposes removing up to 25 people from this workspace, as Funnel's Users page does. They stay in the subscription and keep any other workspaces. Changes nothing: returns change to show the user and a proposal for execute_workspace_user_removal. change lists each email as remove (with its current roles, and keeps_access for a subscription owner or admin, who still reaches the workspace through that role) or not_member. No member listed (status: unchanged) needs no approval. The user cannot remove themselves. Needs get_account and edit_account (Workspace Administrator) on the workspace. Show the user change: each member to remove with their roles, saying which ones keep access through their subscription role, and any email that is not a member. Ask them to approve it; only then call execute_workspace_user_removal with the proposal_id. No one is removed until then.
prepare_rename_workspace
ChatGPTProposes renaming this workspace, as Funnel's workspace settings do. Changes nothing: returns change (the current and the new name) to show the user and a proposal for execute_rename_workspace. A workspace that already has this name is unchanged and needs no approval. Needs get_account and edit_account (Workspace Administrator) on the workspace. Show the user change: the workspace's current name and the new one. Ask them to approve it; only then call execute_rename_workspace with the proposal_id. Nothing is renamed until then.
prepare_restore_budget
ChatGPTProposes restoring an archived budget (an id from list_archived_budgets) and its data source. Changes nothing: returns change to show the user and a proposal for execute_budget_change. not_found when the workspace has no such archived budget. Needs get_account, edit_plan_budgets, edit_fields and edit_datasources. Show the user change and ask them to approve it; only then call execute_budget_change with the proposal_id. Nothing is changed until then.
prepare_restore_convention
ChatGPTPropose restoring an archived convention and its linked fields and adherence. Find its id in list_archived_conventions; review generated_dimensions (before/after) before approval. Nothing is written until execute_convention_change. Read list_guides(guide_id='conventions') first. Show the user the full change, including generated_dimensions before/after and linked-field effects; ask for approval. Only then call execute_convention_change with proposal_id. Nothing is changed until then.
prepare_run_export
ChatGPTProposes running one enabled export now, outside its schedule. Without dates it runs its own date range; with date_start and date_end it reruns that range only. Changes nothing: returns change to show the user and a proposal for execute_export_change. A disabled export is invalid_input. Needs get_account + edit_shares. Show the user change and ask them to approve it; only then call execute_export_change with the proposal_id. Nothing is changed until then.
prepare_set_export_enabled
ChatGPTProposes turning one export's schedule on or off. Changes nothing: reads the export and returns unchanged when it already is in that state, otherwise change to show the user and a proposal for execute_export_change. Needs get_account + edit_shares. Show the user change and ask them to approve it; only then call execute_export_change with the proposal_id. Nothing is changed until then.
prepare_set_portal_dashboard_groups
ChatGPTProposes replacing how a portal's dashboards are grouped and ordered. Groups and the dashboards in each are in display order; name: null is the ungrouped dashboards, and dashboards the portal has that no group lists are appended to it. Every id must be a dashboard in the portal (list_portal_dashboards) and appear once; group names must be unique. before and after are the arrangement as get_portal_dashboard_groups shows it. Changes nothing: returns change to show the user and a proposal for execute_portal_change. not_found when the workspace has no such portal. Needs get_account + edit_portals. Show the user change and ask them to approve it; only then call execute_portal_change with the proposal_id. Nothing is changed until then.
prepare_set_workspace_users
ChatGPTProposes inviting people to this workspace or changing existing members' workspace roles, up to 25 emails, as Funnel's Users page does. New users join the subscription as members and get an invitation email. Changes nothing: returns change to show the user and a proposal for execute_workspace_user_change. change lists each email as invite, role_change (an existing member whose roles are replaced) or unchanged; unchanged alone (status: unchanged) needs no approval, unless roles were omitted where change.roles_editable: then the proposal lets the user pick roles. Needs get_account and edit_account (Workspace Administrator) on the workspace. Show the user change: each email with its roles, and for a role_change the roles it replaces. Ask them to approve it; only then call execute_workspace_user_change with the proposal_id. Nothing changes until then.
prepare_update_brand
ChatGPTProposes renaming a portal brand of the subscription that owns the workspace or changing its colours; every portal using it changes with it, in any workspace of the subscription. Only the fields given change and the logo is kept. Reads the brand and returns unchanged when it already is that way. Changes nothing: returns change to show the user and a proposal for execute_brand_change. not_found when the subscription has no such brand. Needs get_account on the workspace and edit_subscription_portals on its subscription. Show the user change and ask them to approve it; only then call execute_brand_change with the proposal_id. Nothing is changed until then.
prepare_update_budget
ChatGPTProposes renaming an active budget or changing its description. Only the fields given change; description: "" clears it. Targets, dates and breakdowns are edited in Plan, not here. Changes nothing: reads the budget and returns unchanged when it already is that way, otherwise change to show the user and a proposal for execute_budget_change. not_found when the workspace has no such active budget. Needs get_account, edit_plan_budgets, edit_fields and edit_datasources. Show the user change and ask them to approve it; only then call execute_budget_change with the proposal_id. Nothing is changed until then.
prepare_update_convention
ChatGPTPropose replacing an active convention's full definition. Omitted optional fields are removed. Patterns assign each connector to at most one rule; omitted/empty sourceTypes is the all-other-connectors fallback. Read the existing definition first; review full before/after and generated_dimensions, including retained fields that may change. Plan API may create, update or archive many fields. Nothing is written until execute_convention_change. Read list_guides(guide_id='conventions') first. Show the user the full change, including generated_dimensions before/after and linked-field effects; ask for approval. Only then call execute_convention_change with proposal_id. Nothing is changed until then.
prepare_update_portal
ChatGPTProposes renaming a portal, changing its description, or setting, changing or removing its brand. Only the fields given change; brand_id: null and description: null clear them. Reads the portal and returns unchanged when it already is that way. Changes nothing: returns change to show the user and a proposal for execute_portal_change. not_found when the workspace has no such portal. Needs get_account + edit_portals. Show the user change and ask them to approve it; only then call execute_portal_change with the proposal_id. Nothing is changed until then.
prepare_update_scheduled_task
ChatGPTProposes editing an existing scheduled task by id, or enabling/disabling (pausing) it. Provide the schedule_id (from list_scheduled_tasks) plus at least one field to change: name, prompt, schedule_expression, timezone, status ('enabled'|'disabled'), delivery_channels, or slack_channel. Disabling pauses firing without deleting the task. Use rate(N minutes|hours|days) e.g. rate(1 day), or a 6-field AWS cron cron(minute hour day-of-month month day-of-week year) with '?' for an unspecified day, e.g. cron(0 9 ? MON ) for every Monday 09:00. Results go to the workspace members picked as email recipients, by default the task owner. Scheduled tasks can also be sent to other workspace members, but you cannot add, remove, or view those recipients: tell the user to manage them in the Scheduled Tasks page. Updating a task here keeps its existing recipients. Never say a task can only be sent to one person. slack_channel posts Slack results into a channel instead of the owner's DM. Pass the channel name the user gave (e.g. #marketing); it is looked up among the channels the Funnel AI app is in. Never ask for or quote a channel ID: refer to channels by name only. If the channel cannot be used, relay the returned message: the user may need to invite the app with /invite @Funnel AI in that channel, or confirm the name. If it is unclear which channel the user means (e.g. #marketing or #marketing-eu), ask the user to choose before setting it. Only the task owner can change it. When a task delivers to Slack and the user has not named a channel, mention once that results can go to a Slack channel instead of their DM. Changes nothing: returns unchanged when the task already is that way, otherwise change to show the user and a proposal for execute_scheduled_task_change. not_found when the workspace has no such task. Needs get_account and edit_account. Show the user change and ask them to approve it; only then call execute_scheduled_task_change with the proposal_id. Nothing is changed until then.
prepare_update_workspace_settings
ChatGPTProposes changing this workspace's settings, as its About page in Funnel does. The settings are those on the workspace's About page in Funnel: name, and week_starts_on, the first day of the workspace's week, which weekly dates, week numbers and week-based periods follow everywhere in the workspace. Give only the settings to change. Changes nothing: returns change (each changed setting's current and new value) to show the user and a proposal for execute_update_workspace_settings. When every given setting already has that value the answer is unchanged and needs no approval. Needs get_account and edit_account (Workspace Administrator) on the workspace. Show the user change: each changed setting's current and new value. Ask them to approve it; only then call execute_update_workspace_settings with the proposal_id. Nothing is changed until then.
query_data
ChatGPTQueries advertising data from connected platforms. Returns tabular results into a named session table for further analysis. 1. Pick a mode: - query — fetch fresh data by dimensions and measures for a date range - compare_channels — compare metrics across all platforms (cross-platform fields ONLY) - aggregate — re-slice a previously loaded table at different grain (no new fetch) - compute_metric — calculate a ratio (CPC, CTR, ROAS) on a loaded table - restore — reload data evicted from session memory 2. SCOPING RULES (critical — most errors happen here): - Each field is scoped to cross_platform or a specific platform. - Platform-specific fields require source_id (auto-resolves for single-source platforms). - All-cross-platform queries must omit source_id (returns workspace-wide totals). - All platform-specific fields in a query must belong to the same platform. 3. Required fields per mode: - query: workspace_id, result_table_name, dimensions (use [] for totals), measures, date_start, date_end - compare_channels: workspace_id, result_table_name, measures, date_start, date_end (optional: metrics for ratio computation) - aggregate: workspace_id, source_table_name - compute_metric: workspace_id, source_table_name, ratio_id - restore: workspace_id, evicted_table_name 4. Source disambiguation: When multiple sources exist and none is specified, the response returns source_selection_required with available_sources grouped by account. Platform-specific queries accept source_id to resolve. Cross-platform queries accept source_overrides with one source_id per group (a platform may have multiple groups if it has multiple accounts). Dates: YYYY-MM-DD. Percentages as decimals (0.021 = 2.1%). Default limit 500 rows (full result kept server-side; if truncated, re-slice with mode="aggregate" rather than raising limit). Examples: Cross-platform totals: {"mode":"query","workspace_id":"<id>","result_table_name":"totals","dimensions":[],"measures":["common-cost","common-impressions"],"date_start":"2025-01-01","date_end":"2025-03-31"} Single-source with platform fields: {"mode":"query","workspace_id":"<id>","result_table_name":"fb_campaigns","dimensions":["facebookads-campaign_name"],"measures":["common-cost","common-clicks"],"source_id":"<source_id>","date_start":"2025-01-01","date_end":"2025-03-31"} Daily breakdown (re-aggregate): {"mode":"aggregate","source_table_name":"fb_campaigns","group_by":["facebookads-campaign_name","date"],"measures":["common-cost"],"date_granularity":"day"}
reconfigure_data_source
ChatGPTReconfigure an existing source. mode defaults to template; only curated Funnel templates are supported today. Send only template_id, never raw definition or recipe. Inspect the exact UUID first; present workspace, source name/connector, current and proposed template IDs and reconfiguration impact, and obtain separate explicit approval. Connector template listings are candidates, not proof of switch eligibility. If the requested template is already configured, this returns unchanged without writing. Depository checks template availability, connector compatibility and ownership; unsupported connectors and disabled sources are rejected. Pass inspected name, state and template ID as stale preflights, not an atomic lock. One PATCH only; if uncertain inspect, never retry automatically.
reconnect_data_source
ChatGPTReconnect one existing source to a different connection. Inspect the exact source UUID first; present workspace, source name/connector, current and proposed connection IDs and impact, and obtain explicit approval for THIS reconnection. The new connection must be authorized for the actor; Depository checks its access. Pass inspected name, state, and connection ID as stale preflights, not an atomic lock. One PATCH only; if uncertain inspect, never retry automatically. Reconnecting may clear ERROR blockers but does not guarantee healthy downloads.
rename_data_source
ChatGPTRename one exact source by UUID from search_fields category=source only after the user approves the workspace, source, current name, and new name. First inspect it, then pass that current name as expected_name. The private Depository key is resolved internally. A stale name prevents the write. A timed-out/uncertain write must be inspected before any further action; never retry automatically.
request_data_source_access
ChatGPTCreate a Request access link for one exact connector ID, for when the user does not have the login and someone else (for example their client) owns the account. The user sends the link to that person; when they sign in through it, the connection is added for the user in this workspace. not_available means the plan or the connector does not allow it: tell the user the message and do not retry. No data source is created.
restore_data_source
ChatGPTRestore a soft-deleted source by an exact UUID from list_deleted_data_sources. Deleted sources are excluded from search_fields; never guess a UUID. Inspect first (including DELETED state), present workspace, name/connector and impact, then obtain separate explicit approval of THIS restoration. Restoration clears deletedAt but may not resume downloads or reporting or remove an ARCHIVED read blocker. Pass inspected name and state as non-atomic stale preflights; no connection or template is needed. PATCH once; if uncertain inspect, never retry automatically.
search_fields
ChatGPTFinds dimensions, measures, metrics, and connected Data Sources in a workspace. semantic matches natural-language concepts; browse navigates by platform/category/group and browse category=source is the canonical Data Source discovery path; search performs exact keyword matching. Source results return canonical id plus deprecated source_id compatibility, true totals, and truncation. remote_key identifies an external account, not one Data Source. Optional data_source scope resolves the source platform automatically. Returned fields still come from the workspace/platform catalog, not an exact-source schema; catalog_scope=platform makes that evidence boundary explicit. inferred and workspace scope do not narrow field discovery. Use returned canonical id in query_data, prepare_data, or get_dimension_values scope={type:"data_source",id:"..."}.
search_tools
ChatGPTA special tool that returns a trimmed down list of tools given a context. Use this tool only when there are many tools available and you want to get a subset that matches the provided context.
select_data_source_connection
ChatGPTSelect a connection from an immutable setup snapshot. The connection is freshly revalidated, then a new snapshot is returned after remote-account discovery.
select_data_source_remote_account
ChatGPTSelect one remote account from an immutable setup snapshot. The signed handle is freshly revalidated and a new snapshot stores its canonical identity receipt.
select_data_source_template
ChatGPTSelect one curated template from an immutable setup snapshot. The template is freshly revalidated and returned in a new completed setup snapshot. No data source is created.
set_data_source_downloads_enabled
ChatGPTEnable or disable future downloads for one source. First inspect its UUID from search_fields category=source and downloadsEnabled flag; if missing, stop. expected_downloads_enabled must match the inspected value. Disabling does not remove previously downloaded data; enabling does not guarantee healthy downloads. Present workspace, source name/connector, current and desired download state and impact; obtain separate explicit approval of THIS exact action. Use inspected name and flag as stale-write preflights, not an atomic lock. Never change a deleted source. Invoke once; if uncertain, inspect instead of retrying. stale_flag.currentValue is downloads_enabled.
set_data_source_use_in_funnel
ChatGPTControl 'Do not use in Funnel' for one source without stopping downloads. First inspect its UUID from search_fields category=source and useInFunnel flag; expected_use_in_funnel must match the inspected value. Present workspace, source name/connector, current and desired use-in-Funnel state and impact; obtain separate explicit approval of THIS exact action. Inclusion does not guarantee immediate data availability. Use inspected name and flag as stale-write preflights, not an atomic lock. Never change a deleted source. Invoke once; if uncertain, inspect instead of retrying. stale_flag.currentValue is use_in_funnel.
start_data_source_authentication
ChatGPTReturn the safe Funnel browser login link for one exact connector ID. Prefer start_data_source_setup for the complete setup flow.
start_data_source_setup
ChatGPTStart one data source setup using the connector_id from list_connectors. It revalidates the connector and available connections, persists an immutable one-hour EFS snapshot, and may return a Funnel sign-in link. creation_unsupported means this connector cannot be created here: stop and tell the user to finish setup in Connect.
visualize_budget_pacing
ChatGPTShow a Plan budget's pacing as an interactive Funnel view in clients that support MCP Apps: per target (spend first) the actual to date against the period's target and the target to date, with a running spend chart. Takes a budget_id from list_budgets. Reads the current budget period by default, or the period containing date. Values are in the budget's currency. A fixed target (e.g. a target CPC) is compared with the period's value, not paced. Clients without MCP Apps receive the same numbers as text.
visualize_creatives
ChatGPTShow the ads with creative images in Funnel as an interactive image grid in clients that support MCP Apps, ranked by a metric over a date range, with up to 4 metrics per creative. Pass Funnel field ids from search_fields: metric_ids such as common-cost, and optional dimension_ids such as facebookads-ad_name to label each creative (ads sharing the value are grouped). Ads without an image in Funnel are left out. Images load from presigned links that stop working an hour after the call; call again to refresh them. Clients without MCP Apps receive the ranking as text. You receive the ranking but not the images; call get_creative_images to look at them.
visualize_data
ChatGPTRender an interactive Funnel chart (hover tooltips, legend, axes formatted by unit) in clients that support MCP Apps. Prefer source for data already stored by query_data, especially anything over about 30 points: pass workspace_id and source { table_name, x, series: [{ column, name?, unit? }], breakdown? } instead of retyping values. The tool reads stored results; it does not query fresh data. Source rows are aggregated by x (and breakdown when used); ratio metrics listed in the stored table are computed from summed dependencies, not averaged row ratios. Non-additive or pre-aggregated fields may require re-running query_data at the chart grain. For non-table charts, pass exactly one of source or inline categories and series. Source is not supported for table charts; pass their columns and rows inline. One category per x-axis point and one series per measure or per breakdown value, each with exactly one value per category (null for gaps). - line: categories are ISO dates (YYYY-MM-DD). Two series with different units or very different magnitudes (e.g. cost and conversions) get a left and a right axis; otherwise each series is one line on a shared axis (e.g. cost per platform). - bar: categories are dimension values; each series is drawn as its own bar chart. - pie: one series; categories are the slices. - combo: ISO date categories, exactly two series, one bar and one line. Set each series display to bar or line; defaults are first bar, second line. - big_number: one category naming the period and one value per series, displayed as metric tiles. - sparkline: ISO date categories; a compact trend row per series with its latest bucket's value. - table: columns [{name, label, role: "dimension"|"metric", unit?}] and rows of string|number|null cells (up to 500 rows), instead of categories/series. Metric cells must be numbers or null. Optional totals maps metric column names to supplied totals; do not sum non-additive metrics. Tables sort and paginate locally. Copy Markdown uses display formatting; CSV and TSV preserve raw numbers with unit-labelled headers. Set unit "monetary" for money and pass currency, so values format as currency. Pass workspace_id to add an Open in Data Explorer button, and date_range to label the chart with the queried period. Clients without MCP Apps receive a plain-text summary.
x_amz_bedrock_agentcore_search
ChatGPTA special tool that returns a trimmed down list of tools given a context. Use this tool only when there are many tools available and you want to get a subset that matches the provided context.
apply_field_mutation
Claudefunnel___get_dimension_values
Claudefunnel___get_workspace_context
Claudefunnel___list_workspaces
Claudefunnel___load_workspace
Claudefunnel___prepare_data
Claudefunnel___query_data
Claudefunnel___search_fields
Claudeget_dimension_values
Claudeget_workspace_context
Claudelist_guides
Claudelist_workspaces
Claudeload_workspace
Claudemeasurement_check_incrementality_data_readiness
Claudemeasurement_compare_media_plans
Claudemeasurement_create_incrementality_analysis
Claudemeasurement_create_incrementality_test
Claudemeasurement_create_incrementality_test_design
Claudemeasurement_create_media_plan
Claudemeasurement_create_mmm_model_graph
Claudemeasurement_delete_incrementality_analysis
Claudemeasurement_delete_incrementality_test
Claudemeasurement_deploy_mmm_trial
Claudemeasurement_get_forecast_model_accuracy
Claudemeasurement_get_incrementality_actual_vs_counterfactual
Claudemeasurement_get_incrementality_analysis_daily_results
Claudemeasurement_get_incrementality_candidate_power
Claudemeasurement_get_incrementality_daily_lift
Claudemeasurement_get_incrementality_design_candidates
Claudemeasurement_get_incrementality_geo_data
Claudemeasurement_get_incrementality_region_correlation
Claudemeasurement_get_incrementality_region_volume_share
Claudemeasurement_get_incrementality_test
Claudemeasurement_get_incrementality_test_design
Claudemeasurement_get_incrementality_test_preview
Claudemeasurement_get_latest_mmm_deployment
Claudemeasurement_get_media_plan
Claudemeasurement_get_media_plan_channel_allocation
Claudemeasurement_get_media_plan_daily_timeline
Claudemeasurement_get_media_plan_pacing
Claudemeasurement_get_media_plan_projected_results
Claudemeasurement_get_media_plan_run_status
Claudemeasurement_get_mmm_calibration_attribution_data
Claudemeasurement_get_mmm_experiment_config
Claudemeasurement_get_mmm_graph_active_calibrations
Claudemeasurement_get_mmm_input_data_freshness
Claudemeasurement_get_mmm_input_data_statistics
Claudemeasurement_get_mmm_input_timeseries
Claudemeasurement_get_mmm_model_graph
Claudemeasurement_get_mmm_trial_channel_breakdown
Claudemeasurement_get_mmm_trial_fitted_vs_actual
Claudemeasurement_get_mmm_trial_metrics_table
Claudemeasurement_get_performance_report
Claudemeasurement_get_source_saturation_curve
Claudemeasurement_list_accounts_and_partitions
Claudemeasurement_list_forecast_channels
Claudemeasurement_list_incrementality_analyses
Claudemeasurement_list_incrementality_metrics
Claudemeasurement_list_incrementality_regions
Claudemeasurement_list_incrementality_test_designs
Claudemeasurement_list_incrementality_tests
Claudemeasurement_list_media_plans
Claudemeasurement_list_mmm_available_graph_nodes
Claudemeasurement_list_mmm_calibration_options
Claudemeasurement_list_mmm_deployments
Claudemeasurement_list_mmm_experiments
Claudemeasurement_list_mmm_model_graphs
Claudemeasurement_list_performance_channels_and_sources
Claudemeasurement_replace_mmm_graph_calibration_config
Claudemeasurement_run_forecast_what_if
Claudemeasurement_run_media_plan
Claudemeasurement_start_mmm_experiment
Claudemeasurement_update_media_plan
Claudemeasurement_update_mmm_model_graph
Claudemutate_fields
Claudeprepare_data
Claudequery_data
Claudesearch_fields
Claudesearch_tools
Claudex_amz_bedrock_agentcore_search
Claude