MCP App Store
by KeenEthics
PipeLedger icon

PipeLedger MCP Server for ChatGPT

By PipeLedger Inc.

Overview

PipeLedger is the financial intelligence layer for the agentic enterprise. We help businesses put their financial data to work with AI while keeping control of who can access it and what they can do. PipeLedger connects QuickBooks Online and NetSuite, standardizes financial data across systems, and adds the accounting definitions and business context agents need to work with the numbers. In ChatGPT, explore revenue, cash flow, company balances and operating activity. Compare results across periods and entities, investigate changes, trace figures to supporting transactions, and create financial statements and reusable reports. Your finance team controls each connection’s access and permissions. Identity masking protects sensitive names while preserving useful analysis. Separately authorized connections can maintain reporting definitions, record unit activity and request data refreshes. PipeLedger tools do not post transactions to your accounting system or initiate payments.

How to connect

  1. ChatGPT: Open the PipeLedger listing in the ChatGPT app directory and choose Connect. ChatGPT listing PipeLedger exposes 16 tools here.

Tools

pl_audit

ChatGPT
Read the immutable audit trail for this org. Use when: the user asks who changed what and when, wants to know which runs were approved / rejected, or is investigating a pipeline, mart-lifecycle, or credential-governance event. Results are scoped to the caller's org automatically and limited to the curated governance-action vocabulary advertised by actions. Do NOT use when: the user wants GL transactions (use pl_query against gl_lines) or a financial report (use pl_report). audit_logs are metadata about pipeline + approval events, not financial data. Filter by exact action, governance domain, structured actor, accountable human, resource, pipeline run, or credential. Returned events contain immutable identifiers plus allowlisted actor/resource/access references, mart labels, constrained mart risk classes, and timestamps when recorded. Untyped resource row-count values are intentionally excluded; summary reference coverage may still show how many audit events recorded row-count evidence. actor_context_status distinguishes complete or partial runtime actor evidence, responsible-user-only evidence, and missing actor context. Raw changes/details payloads and IP/network context are intentionally excluded from the ChatGPT surface because they can contain unnecessary security metadata. No input flag can override that boundary. Action names ending in _changed prove only that a configuration mutation occurred. They do not reveal its direction, affected values, whether access expanded or contracted, or the resource's current state. Do not infer those facts from the action name or from the absence of a later event. For broad or large-population audit reviews, call response_mode=summary first across one fixed window. Use its exact action, actor-context, reference, resource, pipeline-run, and constrained risk-class counts to identify coverage gaps and choose narrow follow-up filters. Then call response_mode=events for only the relevant actions, actors, resources, or pipeline runs, normally 50-100 events. Do not exhaustively page or assemble a workbook merely to calculate counts already returned by summary. Summary is aggregate evidence: action totals alone do not prove per-resource lifecycle ordering or complete certified → approved → published chains. If the user explicitly requests a complete row-level ledger, use limit=250 and follow next_cursor with every filter unchanged. Otherwise disclose that cross-event chain validation requires narrowed event evidence or a dedicated deterministic server-side review/export; do not invent chain findings from aggregate counts. Window is capped at 90 days per request. Event pagination is stable and filter-bound. Move start_time/end_time only to investigate a different window. Privacy and billing: this call records a private PipeLedger usage event containing the organization-scoped credential, tool, status, timing, rows, and PipeLedger compute/warehouse units. That usage event does not store the host conversation, prompt, assistant messages, generated SQL, or returned financial rows. Query reconciliation uses structural metadata and cryptographic fingerprints rather than readable filter or scope values.

pl_catalog

ChatGPT
Read governed numeric inputs with action=assumptions and assumptions.as_of; these are controller-supplied values, not calculated metrics. Read the org's finance metric catalog — the governed registry of metric definitions that pl_query and pl_report resolve metric IDs against. action='list' (default) returns the effective catalog: canonical finance metrics PLUS every org-authored definition (draft / active / archived) with status and change-note comments, plus active saved Metrics Reports and recipe summaries, each labeled by definition_kind. Load recipe IDs with pl_report action=load_recipe; only Metrics Reports expose an executable metric_ids packet. Use it to discover metric_ids, stable report_ids, and see what's pending vs live before authoring (pl_catalog_admin) or before passing metric IDs or a report ID to pl_query or pl_report. A temporal_ratio definition shows its numerator and denominator operands and fixed temporal contract, and intentionally has no singular source_mart; execute it through pl_report report_type=metric_report, not pl_query. Internal database, organization, and actor identifiers are omitted. Pass catalog_version to list a supported catalog version; omit it to use the current supported version. action='get' returns one definition. Pass metric_id (e.g. 'opex') for the sealed definition resolved through the effective catalog — including definition_hash and catalog_version — without fetching the whole catalog; optional catalog_version selects a supported catalog version (the current supported version when omitted). action='history' returns the newest-first revision changelog for one organization-authored metric_id. Use its stable metric_revision values when the user asks how a prior definition would change the current report; internal definition-row UUIDs are not returned. action='preview' compiles the effective catalog with an inline unsaved definition overlaid (the same dry-run the Schema UI shows pre-activation) so you can read the resulting effective_catalog_hash and sealed metric WITHOUT activating it. action='versions' returns registered release metadata in version order; use it to discover available versions. action='compare' compares two registered system catalog versions for this organization. Both version values must appear in action='versions'; never infer an unlisted version. It explains added, removed, and changed metrics, whether calculation logic changed, definition hashes, effective hashes, and system-versus-organization definition sources. Cite hashes in short form (first 12 characters plus an ellipsis) in prose; the complete value stays in structured content. It does not calculate financial amounts; use pl_report definition_comparison for the isolated numerical impact of one organization-authored metric revision. action='governance' returns one agent-oriented snapshot that binds—but does not merge—the current Finance Catalog SemVer/hash and organization account-semantics revision. It also returns bounded newest-first history derived from the semantic clock and immutable audit evidence. Viewer output includes only controller-approved organization comments and safe revision/status evidence; privileged technical notes, actor/account IDs, and raw before/after values are excluded. Publication evidence remains per mart. Does not modify catalog state. Requires only pl_catalog in allowed_tools. Authoring lives in pl_catalog_admin. Privacy and billing: this call records a private PipeLedger usage event containing the organization-scoped credential, tool, status, timing, rows, and PipeLedger compute/warehouse units. That usage event does not store the host conversation, prompt, assistant messages, generated SQL, or returned financial rows. Query reconciliation uses structural metadata and cryptographic fingerprints rather than readable filter or scope values.

pl_catalog_admin

ChatGPT
List, create, edit, delete, activate, and archive governed catalog resources through shared controller-grade lifecycle services. Existing Finance Metric and classification resources use the same governance as Settings → Schema; cash flow subcategory rule authoring is MCP-only in V1. resource='metric_definition' (default): action='create' stores a NEW draft (status='draft'; not live until activated). action='edit' replaces a draft in place, or forks a new draft revision when the target is ACTIVE (the active definition stays live until the revision is activated). action='activate' promotes a draft to active (supersedes the prior active revision of the same metric_id). action='archive' retires a draft/active definition (it leaves the effective catalog). resource='cash_flow_subcategory_rules': the organization's cash flow subcategory rulebook. action='list' returns every rule in ascending precedence — the exact evaluation order. action='create' adds one rule; action='edit' replaces the rule addressed by its stable rule_key; action='delete' removes it. There is NO draft or activation step: a written rule is governed configuration immediately. A rule derives the presentation subcategory of an already-classified indirect cash-flow component from a closed evidence list, and never changes standard cash-flow classification or authoritative amounts. Exactly one matching rule assigns its subcategory; several matching rules resolve to the lowest precedence; no match leaves the component in the synthesized "Unassigned <Section>" bucket. An account-level exception is just a higher-precedence rule keyed on source_account. resource='account_classification': action='list' finds source accounts by name or ERP account id and returns each one's stable classification_id with its current taxonomy and cash-flow classification. action='edit' updates one source account's taxonomy override intent, including Catalog Category/Subcategory (taxonomy_level_4_catalog / taxonomy_level_5_catalog_subcategory) and cash-flow Cash flow category (cash_flow_category), Calculation basis (cash_flow_treatment), and Standard detail (cash_flow_standard_detail), or independently resets the cash-flow override. Pass the stable classification_id returned by action='list' as id; never pass the ERP account id or override_id. The same account-classification governance helper used by the UI performs the write. resource='transaction_override': generic transaction-detail override authoring, selected by transaction_override.domain (currently 'cash_flow_classification', which reclassifies the Cash flow category, Calculation basis, and Standard detail of ONE posted Balance Sheet counterpart GL line). action='list_candidates' returns every override-eligible line with a ready-to-submit payload template, from EITHER transaction_override.verification_item_id (a published Data Review finding) OR a transaction lookup in transaction_override.payload (gl_line_id, cash_flow_line_identity, or transaction_id, optionally narrowed by source_erp, connector_id, transaction_type). A review finding is NOT required. Find the line first with pl_query on gl_lines and pass its gl_line_id unchanged; the service resolves it to the durable cash_flow_line_identity. transaction_id is a search key, never unique: present every returned line, let the accountable human choose, never pick the first match, and read the structured 'truncated' flag rather than prose. Copy a template, edit only the classification fields and the required 10-500 character override_reason, and submit it as transaction_override.payload with action='edit'. Every template carries expected_override_version (the override_version observed on the line, null when none existed); the write is atomic against exactly that observation and answers E_TRANSACTION_OVERRIDE_CHANGED otherwise, so reload the candidates instead of retrying blindly. action='void' removes one existing override using a void_payload template (from list_candidates, or from the transaction-override discovery list on REST/CLI after the originating item resolves); the compare-a…

pl_data_quality

ChatGPT
Checks whether enabled marts are ready, awaiting approval, blocked, or not yet materialized. Omit mart for an all-enabled-mart overview with no mart-table scan; provide mart for one scoped diagnostic scan when that mart has specific signals. diagnostic_execution reports whether that scan ran, its query count, and bytes processed. direct_query_supported describes direct table-query support; it is not a general availability flag. The Business Object Directory remains available through governed object search and listing even though direct table queries are disabled. If a Directory object kind reports supported=false because of an identity-tokenization policy, that is a plaintext name-search limitation for this credential, not a readiness or data-product limitation: the directory still lists that kind with token labels, and those identities are still delivered as deterministic tokens. primary_access_tool is structured routing guidance for the agent; do not surface internal tool names in an ordinary finance-user answer unless technical routing was requested. Use after a pipeline run, during onboarding, when a number is questioned, or before a close/reporting workflow. A failed newer delivery is reported as blocked even when an older publication remains available. A certified newer version awaiting review is pending_approval. refresh_in_progress is unfinished work reported separately; it does not by itself mean a newer certified or unpublished version exists. Quality certifies that PipeLedger’s deterministic logic executed and passed its declared checks; it does not validate source accounting or configuration choices. Privacy and billing: this call records a private PipeLedger usage event containing the organization-scoped credential, tool, status, timing, rows, and PipeLedger compute/warehouse units. That usage event does not store the host conversation, prompt, assistant messages, generated SQL, or returned financial rows. Query reconciliation uses structural metadata and cryptographic fingerprints rather than readable filter or scope values.

pl_drilldown

ChatGPT
Verify an opaque authenticated drilldown handle and return underlying supporting rows or aggregate the same handle-bound population by one authorized dimension. Use when: the user (or you) wants to see what's behind a number on a pl_report line or an aggregate pl_query result. Pass the opaque handle string exactly as received and choose the accounting evidence grain the user asked for. Omit group_by for supporting rows. To answer a bounded question such as "capital expenditures by project", pass one concise pl_schema column where groupable=true, for example group_by=["project_name"]. The grouped call remains bound to the handle's metric, filters, period, publication, and access basis; it never opens a new free-form query. Every issued handle arrives with a sibling supported_target_grains list from pl_report or pl_query; pass one of those values as target_grain. For a Balance Sheet value, target_grain=account_balance reads governed Trial Balance rows only. target_grain=period_activity returns governed account-period Trial Balance rows and a cross-mart bridge. target_grain=transaction_line returns governed GL lines and the same bridge. The bridge is opening Trial Balance + source GL activity + modeled/non-source Trial Balance adjustment = ending Trial Balance; a modeled adjustment, including virtual close, is never presented as a transaction. Derived totals and ratios first return callable component handles, each advertising its own supported_target_grains. Returned shape: - handle_label: human label echoed from the handle (e.g. "Revenue, FY2026"). - catalog_snapshot_verified: true confirms the pinned catalog version, organization-definition rows, effective hash, and metric definition hash reconstructed exactly. The call fails closed if they do not. - newer_catalog_version_available: true means a newer registered system catalog version exists. Supporting rows still use the handle's exact pinned definition. - target_grain: the selected account_balance, period_activity, or transaction_line route. - staleness: compares the handle run with the current reported-value publication and, for cross-mart Balance Sheet drilldowns, current_supporting_published_run_id. Cross-mart reads fail closed unless both marts match the same approved pipeline run, release channel, dataset, and request snapshot. Surface 'stale' to the user — the underlying number may have moved since the report was generated. - result.mode: "rows" for atomic metrics (returns governed rows ordered by absolute authoritative amount DESC with stable ties), "dimension_aggregate" when group_by is present (returns the largest absolute groups plus an exact full-population total). Row mode's returned_row_count counts only rows in this response. Dimension mode distinguishes returned_group_count from authorized_group_count, the complete policy-authorized grouped population; truncated is true when those differ. result.supporting_detail_complete=false means protected account rows were omitted; the rows that remain keep their truthful returned_grain, and reconciliation is withheld. Derived/ratio components include callable component drilldown handles using the same pinned catalog snapshot. - source_provenance.publication_age_seconds is elapsed time since the governed publication, not source-system freshness, accounting completion, or close status. source_data_recency states the separate bounded source-currentness limitation. - result.reconciliation: compares supporting rows to the bound reported value only when the access basis, catalog, and publication still match. A not_comparable result is expected when any of those changed; do not present it as a financial discrepancy. Auth: handles prove provenance, not permission. Your dimension scope, delivery filters, clearance, account confidentiality, and fixed privacy still apply — you may see fewer rows than the subject who originally generated the report. Common failure modes: expired (24h TTL), authentication_failed (reference was altered or is unavailable), unknown_kid (rotated/unsupported key), schema (malformed reference), storag…

pl_post_unit_voucher

ChatGPT
Load, post, update, or void manual unit vouchers (signed non-financial movement entries such as subscriptions, usage, seats, and volumes). The historical name is retained for credential compatibility, but the action field is the operation authority. BEFORE post or update: call pl_unit_register action='list_metrics' — every entry's metric_key must be an ACTIVE, direct-entry metric in the org's registry, and each metric's behavior (balance/flow) constrains which movements it accepts. Metric keys are org-defined; never guess them. SAFE UPDATE: call action='load' by voucher_id or source_ref, preserve every returned entry unless the user explicitly changes it, and copy version verbatim into action='update'. Update replaces the editable header and ALL entries and is rejected when the optimistic-lock version is stale or either the current or requested period is locked. source_ref, correction linkage, and client_metadata are immutable. action='void' is a soft void: entries and audit evidence are retained, but the source_ref is released and published marts exclude the voucher after the next transform. Update and void require an accountable owner/admin. source_ref is the RETRY-SAFE IDEMPOTENCY KEY: resending the exact same payload with the same source_ref returns the original voucher (idempotency_status='replayed'); a reused source_ref with a different payload is rejected as a conflict. Corrections (voucher_type='correction') require corrects_voucher_id or correction_reason and admit only 'adjustment' movements. Readers can validate through pl_unit_register action='validate_voucher'. Set validate_only=true for an advisory dry run that reports the violations it can see at once without posting (registry membership, movement grammar, signs, period locks, dimension values, legal-entity key shape). One gap is post-only: the dry run cannot compare legal_entity_source_key against a STORED voucher, so a replayed source_ref whose key diverges from the original surfaces IDEMPOTENCY_CONFLICT only at post time. A successful post is NOT immediately visible in the unit marts (publication_status='pending_transform'): entries flow into unit_movements / unit_rollforward on the next successful transform + publication, so do not treat a pl_query miss as a lost posting, and never trigger pl_run as a side effect of posting. ATOMIC UNIT METRIC MANAGEMENT (SB306): action='create_metric' | 'update_metric' | 'activate_metric' | 'archive_metric' with the payload in metric. These need the separate allow_unit_metric_management grant (SB307), which an admin enables under Unit Register posting; posting alone does not cover them. create_metric makes a custom metric postable at once; omit metric.movement_labels and the organization's industry proposes the movement names (real estate: new→Started, churn→Sold; lending: new→Originated, churn→Repaid), or pass an explicit map ({} keeps the global names). update_metric changes PRESENTATION only: metric_label, description, metric_category, movement_labels (the map REPLACES the stored one). Grammar (behavior, activity model, unit) is not editable; the activity model freezes once entries exist. Name changes reach the published unit marts on the next successful transform. The accountable human must be an owner/admin, and the credential must be UNRESTRICTED (no dimension scope, no row filters): a scoped credential is refused because registry changes apply organization-wide. Disabled by default — an admin must enable the 'Unit Register posting' grant (pl_post_unit_voucher in allowed_tools AND allow_unit_posting=true) AND set credential_role='operator'; metric actions additionally need allow_unit_metric_management=true. Scope-constrained credentials may only post entries stamped inside their allowed scope and may load/update/void only vouchers wholly inside that scope. On E_TOOL_NOT_ALLOWED, tell the user their credential lacks permission; do not attempt a workaround. Privacy and billing: this call records a private PipeLedger usage event containing the organization-scoped credential, tool, status, timing, rows, and PipeLe…

pl_profile

ChatGPT
Identify the account represented by this connection for host profile labels. Returns a stable opaque profile ID and available name, email and workspace/connection label. The same account keeps its ID across token refresh and reconnect; labels do not identify the calling host. Use pl_whoami for connection instance and permissions. Free authentication-only read; no financial data or FIQ charge.

pl_query

ChatGPT
Run a parameterized query against a certified, published mart. Two modes: When the user's need is analytical (totals, movement, composition, comparison), prefer one aggregation-mode call over enumerating rows; row-mode enumeration of a large population is for explicit export-style requests. When a row population exceeds the page cap and the need reads analytical, offer aggregation before exhaustive paging — each page is a separately metered operation. Row mode (default — omit metrics): returns rows at the selected mart's native grain. For example, gl_lines returns one governed General Ledger line while trial_balance returns an account-period balance row. Use gl_lines when the user asks for transactions; use trial_balance for account-level point-in-time balances. order_by + order_direction sort the rows by any policy-visible mart column (for example reporting_amount DESC for largest lines first), with the mart's stable row keys appended as tiebreakers; omit them for the default stable ordering. Honors dimension scope, GL-account confidentiality, fixed identity privacy, and the subject's projection policy. Choose the starting capability from the strongest available clue; there is no mandatory global order between pl_resolve and pl_query. Use pl_resolve first when the clue is a name, exact source or reporting-object ID, governed or advisory role, corporate family, identity, parent, or other known registry relationship. Use pl_schema and pl_query first when the clue is a transaction, memo, item or service, document reference, purchase order, or other financial activity. Compose the capabilities in either direction as needed, preserving exact returned object types and IDs. When the starting clue is financial activity, determine the intended governed object_type; if it is ambiguous, ask for clarification or keep separate object-type paths distinct. pl_resolve does not search transaction rows, and pl_query activity never creates a role, identity link, same-kind consolidation, corporate-family membership, or cross-register equivalence. Inspect gl_lines with pl_schema, then use row-mode pl_query with contains_ci to search each available, policy-permitted evidence field separately because filters combine with AND. Typical evidence fields are memo, item_name, document_reference, po_number. Use statement_type="all" unless the user explicitly narrowed the financial statement. Select only the matched evidence field and the requested object's ID/name pair: customer → customer_id and customer_name; vendor → vendor_id and vendor_name; employee → employee_id and employee_name; project → project_id and project_name; legal_entity → reporting_legal_entity_id and legal_entity_name. Deduplicate the requested object ID and preserve multiple matches. When a bounded candidate set is already available from pl_resolve, use schema-supported multi-ID filters and grouping to query it in batches; keep Customer-side sales/receivables evidence separate from Vendor-side bills/payables evidence. Never reuse a Customer ID as a Vendor ID or join records by a similar name. Call pl_resolve with the same object_type and each exact returned ID when governed registry provenance is needed. Across every capability, prefer one page-, filter-, or in-batched call over any per-record loop; call volume is metered per operation. An activity-evidence match proves only a historical row association; it does not prove current capability, identity, business role, or causation. Memos, item names, references, document identifiers, and PO numbers remain activity evidence—not resolver object types or Business Object Registry entries. If an evidence or object field is unavailable or rejected as masked, do not probe it through another field or tool. Project route: for Real Estate-specific project health, development economics, financing, LTC, or equity questions, prefer project_overview_real_estate when the token allows it. The general project_overview remains queryable and is the preferred surface for broader operational measures such as project manager, status, hours, b…

pl_report

ChatGPT
Use action='load_recipe' with report_id and report_library_scope to open a saved report recipe discovered through pl_catalog. It returns a recipe definition with author-provided presentation notes, not financial values. Follow its steps using ordinary tool calls under the reader's permissions. Omit report_type and financial arguments when loading. Generate an Income Statement, Balance Sheet, Cash Flow Statement, or configurable Metrics Report from approved published marts. Set exclude_intercompany=true to exclude controller-tagged intercompany accounts from Income Statement, Balance Sheet, Cash Flow Statement, or ordinary Metrics Report metrics sourced only from GL lines, Trial Balance, and Cash Flow Components. The same exclusion applies to comparisons and drilldowns. An excluded Cash Flow Statement is a partial ledger view that retains its calculated reconciliation difference. Mapping or allocation gaps remain unreconciled; otherwise it is coverage-qualified. Temporal ratios and other source marts reject this option. Exclusion does not match or eliminate transaction pairs, and untagged does not mean confirmed third-party. Excluded Metrics Reports do not issue save handles because saved definitions do not retain this option. Use only these canonical PipeLedger report-family names: Income Statement, Balance Sheet, Cash Flow Statement, and Metrics Report. Only the first three are financial statements. P&L and profit-and-loss are user-intent aliases for Income Statement; they are not report_type values. Use Metrics Report as the sole name for a configurable governed-metric report. - Use pl_query for focused account queries or supported metric aggregations, such as Revenue or Cost of Goods Sold totals, on the appropriate governed mart. A full Income Statement additionally returns statement lines and accounting controls; use pl_report when those are requested. Catalog, readiness and governance/audit evidence support either route. A governed report value does not require a duplicate query merely to be auditable. - KPI, management report, management reporting, executive report, board report, board pack, dashboard, and cross-mart report are ambiguous presentation requests, not report aliases, report_type values, saved templates, or default Metrics Report definitions. Do not map those words alone to Metrics Report. They may title an overall user-requested work product such as "Monthly Board Pack", but every embedded PipeLedger report keeps its canonical Income Statement, Balance Sheet, Cash Flow Statement, or Metrics Report title. - A Metrics Report requires the user's explicit ordered metric selection. If the user has not supplied every desired metric and its order, do not choose a standard packet or invent metric_ids. Ask for the ordered metric list. If the user delegates selection, propose an ordered list and wait for confirmation before resolving metrics or calling pl_report. - Never infer a balance as_of endpoint from a fiscal year, current date, latest publication, activity coverage, or available data. After the selected metric grains are known, ask for each missing current or comparison clock required by Metrics Report. A fiscal_year alone does not authorize a balance endpoint. - Sub-year windows route by clock, not tool preference. A Balance Sheet or balance-grain metric for a specific month stays in pl_report: any valid as_of date selects that calendar month's authoritative month-end position, so a mid-year position is never a reason to switch tools. An Income Statement view for a single month, quarter, or other sub-year activity window has no pl_report clock: answer it with one aggregation-mode pl_query over that statement's governed metric_ids on mart=gl_lines, filtered by posting_period or fiscal_quarter_key, and present it as a governed period view rather than a complete or certified statement, offering the fiscal-year Income Statement separately. A sub-year Cash Flow Statement view is not a governed route: offer the fiscal-year Cash Flow Statement, and use mart=cash_flow_components only for month-grain component an…

pl_report_library

ChatGPT
Save, revise, archive, propose, or withdraw reusable Metrics Reports and report recipes. Recipes accept a validated recipe object instead of a save_handle; this stores intent and executes no financial steps. Open recipes with pl_report action='load_recipe'. For Metrics Reports, first run pl_report, then pass its signed save_handle here. The handle preserves the exact ordered metric and section placement that was successfully executed. Metric result values and resolved execution dates are not stored in Metrics Report definitions; when the user explicitly confirmed one, the handle may include a reusable relative-period policy. library_scope='personal' writes only to the active accountable human's private library. library_scope='organization' requires credential_role='operator' and allow_organization_report_write_access=true in addition to this tool being granted. Saves and revisions become active immediately when those gates pass. Save/revise of an identical active definition and archive of an already archived definition can return idempotent_replay=true. Propose can replay an existing pending proposal. These are definition/proposal outcomes; every invocation still records its own usage event. If you lack organization write access, save to the personal library and use action='propose'. That submits the report to the organization's owners and admins for review; it does not publish anything and needs no extra grant. A reviewer chooses the organization report ID when they approve, so you do not pick one. Use action='withdraw' to retract a proposal you have not had decided yet, for example to propose a newer revision. A successful withdrawal returns promotion_request.state=withdrawn, confirming completion. Withdrawal acts on a pending proposal; it is not a status lookup or an idempotent replay. If no proposal remains pending, it returns E_NOT_FOUND; that error alone does not establish whether a prior withdrawal succeeded. Saved definitions are presentation contracts, not data grants. Every run re-resolves referenced metric IDs through the current active Finance Catalog and applies the executing credential's current delivery policy. Privacy and billing: this call records a private PipeLedger usage event containing the organization-scoped credential, tool, status, timing, rows, and PipeLedger compute/warehouse units. That usage event does not store the host conversation, prompt, assistant messages, generated SQL, or returned financial rows. Query reconciliation uses structural metadata and cryptographic fingerprints rather than readable filter or scope values.

pl_resolve

ChatGPT
action=context retrieves published effective-dated legal profiles, reporting groups and direct ownership interests. Use object_type=legal_entity and context with exactly one as_of date or inclusive period; omit query. Batch exact business_identity_ids, reporting_legal_entity_ids or party_references for forward lookup; group_ids or owner_identity_ids select companies in reverse and return permitted Customer/Vendor reporting IDs for pl_query. Preserve kind, ERP, connector and dated intervals; never name-match or apply today's ownership to old postings. Follow next_cursor with unchanged selectors. This mode requires explicit Entity Context and Business Object Directory access, Highly Restricted clearance and permitted original identities; it never grants access or changes financial amounts. Ownership shares are context, not control, tax allocation or consolidation instructions. PipeLedger's governed semantic resolver for finance metrics and business objects. action=search is the primary mode: it translates a natural-language label, business term, exact durable ID, source binding, alias, or permitted corporate-family name into policy-visible candidates, with deterministic search confidence, ambiguity guidance, durable identity authority, advisory ERP evidence, relationships, source provenance, and publication evidence. Search requires query and exactly one object_type: metric, customer, vendor, employee, project, or legal_entity. action=list is an additional directory mode for customer, vendor, employee, project, or legal_entity. It omits query and enumerates one governed, policy-visible registry at reporting-object grain; use pl_catalog to list metrics. directory_count reports the exact authorized reporting-object population before the response limit, the returned count, and whether the response was truncated. Directory names are source-qualified for readability while exact ERP IDs remain provenance and reporting_object_id remains the governed same-kind filter value. source_binding_count covers only the exact source bindings visible under the caller's policy for that returned reporting object. Exact name-source provenance is returned only when that source binding is also policy-visible. Under an identity-tokenization policy the directory is still listed: display_label_kind=token means each label is the same deterministic identity token pl_query, pl_report, and pl_drilldown deliver for that object under this credential, identity_tokens lists every token the object presents across its policy-visible connectors, and reporting_object_id_kind states whether the governed filter ID or the token itself is the returned reference. Preserve tokens exactly; they are the cross-tool reference for that object. action=review is the exhaustive, read-only evidence-review mode for customer, vendor, or employee. It returns one exact source record per review_entry, not a ranked reporting-object search result. query is optional. review_filters can narrow availability, lifetime activity evidence, governed role/state, advisory role/state, identity state, source ERP, and stable evidence_basis. Values within one filter array are OR; populated filter groups combine with AND. review_page reports the exact filtered policy-visible population and next_cursor. Pagination uses next_cursor with the same object_type, query, and filters; an absent next_cursor marks the end of the filtered policy-visible register. Each page contains the source-record evidence for that page, so per-record searches are not required for register coverage. Each additional call is separately metered. Advisory roles remain distinct from governed roles; year-specific activity, 1099 status, Customer type, and transaction-type evidence are available only where explicit returned fields provide them. This mode creates no assertion, identity, or identity link. It is unavailable under identity tokenization because exact source-record evidence cannot be presented without creating a name/ID oracle; action=list remains available with tokens. The Business Object Directory is intentionally acce…

pl_run

ChatGPT
BI actions: bi_preflight previews existing mart field additions across configured BI sources without warehouse queries; bi_publish refreshes BI from approved data, optionally with add_fields; bi_status tracks a BI request. Field additions require the explicit BI grant, not higher reading clearance. Fields apply organization-wide; existing BI masking and confidentiality remain in force. Metric delivery stays with Finance Catalog activation. Preflight estimates impact; deployment, release, privacy or catalog changes may rebuild additional sources. Field edits coalesce for three minutes; plain refresh advances pending requests immediately. Cron pickup and completion are asynchronous. The request charges nothing; publication work is metered in FCU by the BI worker. Manage pipeline operations with an explicit action: list discovers pipeline IDs, organization-enabled marts, and recent/in-flight runs; preflight checks a proposed request without queuing; status reads a persisted run; trigger queues work. All actions require Operator and pl_run access. Read actions use FIQ; trigger uses pipeline FCU. For a new request, use list then preflight; audit-history searches are unnecessary. Preflight is an observation, not a reservation; trigger checks current state again. Transform uses data already in PipeLedger without ERP extraction. Full extracts then transforms; it does not imply full_refresh. Incremental is normal extraction depth; use deeper modes only for an identified need. Marts are organization-level delivery selections; the worker resolves build dependencies. Runs may coalesce. Compare effective_request and effective_run; joined runs can have broader marts, force, or automatic publication. For an explicitly bounded transform use bounded=true; incompatible joins are refused atomically, including later joins. Only trigger and bi_publish mutate. Supply a reason and inspect the returned persisted run settings; queued is not completed or published. Do not trigger as a side effect of reporting. A pending inventory valuation requires transform, force=true, authorized auto_approve=true, incremental, and all enabled marts using its isolated workflow.

pl_schema

ChatGPT
Inspect the schema of a published mart before writing pl_query filters or naming metrics. Use when: starting a new question against a mart for the first time in a conversation, or when an earlier pl_query returned an unexpected column shape. Always call pl_schema before fabricating column names — this is the anti-hallucination surface. Response size: detail: "concise" is the default and returns per column only name, type, static filterable/groupable capability flags, column_group heading, and the masking marker — a fraction of the full payload. The load-bearing mart caveats and fiscal_calendar rules stay identical. columns: [...] narrows the response to exactly those columns; pair it with detail: "full" to read complete descriptions for just those fields. Unknown names get close-match suggestions. Returns: - aggregation_supported: whether this mart accepts pl_query aggregation mode. When false, every column is non-groupable and only row mode is available. - columns: concise structural entries by default; full entries preserve the same filterable, groupable, column_group, and masking fields and add nullability and description. Columns with an explicit semantic operator profile also carry filter_kind and mode-separated filter_operators (row vs aggregation); use those to choose predicate shapes instead of assuming one universal grammar — scalar_equality means a bare scalar value compiling to equality. Columns without the two keys keep their existing operator behavior. Capability flags describe the static mart-and-policy contract, while query-specific scope, row filters, account confidentiality, metric grain, and column combinations can still reject a request. Masked columns carry filter_behavior: "refused" means the mask is org-wide/fixed and any predicate on the column errors; "confidential_rows_omitted" means the mask is account-conditional, so the column IS filterable — the protected accounts' rows are omitted from the population and the result carries a confidential_rows_omitted note. A tokenize-protected identity can be groupable into stable private-token buckets regardless of its filter behavior; masked_by_governance never means its original value is visible. - hierarchy columns: every L1-L5 column physically supported by this mart is discoverable when included by the ordinary delivery-column policy. Use _depth for uncapped source-path depth, _max_native_depth to distinguish source-native levels from leaf-propagated levels within L1-L5, and taxonomy_assigned_depth for assigned reporting-taxonomy depth. - recommended_metrics: catalog metrics executable by pl_query whose source_mart matches this mart. Temporal ratios are intentionally excluded because they combine a governed activity window with balance snapshot semantics and have no singular query mart; discover them through pl_catalog/pl_resolve and execute them with pl_report report_type=metric_report. Pass listed metric_id values into pl_query or pl_report rather than reinventing financial statement sign handling. - allowed_time_buckets: semantic period names this mart supports (trailing_12_months/TTM, ytd, current_quarter, ...). Use these instead of computing date ranges by hand. - period_dimensions.native: period columns you can filter/group_by directly. On gl_lines these include fiscal_quarter / trimester (SEASONALITY buckets — combine across years), fiscal_quarter_key / trimester_key (a SPECIFIC closed period like FY2026-Q2), and business_days_in_month / business_days_in_fiscal_quarter (period attributes for business-day normalization — project, never SUM). period_dimensions.via_dim_date_join_future lists period columns not yet on this mart. - fiscal_calendar: the org's fiscal-year anchor (fiscal_year_start_month, fiscal_year_end_month, label_convention). READ THIS before interpreting any period key: the FY label is the calendar year the fiscal year ENDS, so a June-30 year-end org's FY2026 = Jul 2025-Jun 2026 and FY2026-Q4 = Apr-Jun 2026 (not calendar Q4). Don't assume calendar quarters; confirm a key's window from its fiscal_quarter_start_date / fiscal_…

pl_unit_register

ChatGPT
Read active Unit Register metrics, direct-entry eligibility, movement labels and period locks. validate_voucher checks a proposed voucher using Unit Register read access, without posting permission. Its validation.valid describes the payload; posting_eligibility separately describes current write permission. Checks include movements, quantities, dimensions, Legal Entity, period locks and source-reference conflicts. An actual post rechecks permissions and all posting rules. Validation does not reserve a source reference. Metric vocabulary and period locks are organization-level metadata. Voucher validation enforces the connection's dimension scope and supported row filters; filters requiring warehouse-derived fields cannot be bypassed. ERP source bindings require unrestricted unit metric management access. This tool never posts, updates or voids vouchers. Published quantities remain behind pl_query. Privacy and billing: this call records a private PipeLedger usage event containing the organization-scoped credential, tool, status, timing, rows, and PipeLedger compute/warehouse units. That usage event does not store the host conversation, prompt, assistant messages, generated SQL, or returned financial rows. Query reconciliation uses structural metadata and cryptographic fingerprints rather than readable filter or scope values.

pl_usage_evidence

ChatGPT
Read request-level PipeLedger usage evidence for this credential. Use after a multi-tool analysis to correlate opaque host_session_id and request_id values with data-warehouse billed bytes, versioned FIQ units, and billing-record delivery state. By default the tool uses the current host session when ChatGPT or another MCP host supplies session metadata. Pass evaluation_scope=agent with a bounded start_time/end_time window for an exact aggregate across every session owned by this credential; request details remain capped by limit and truncated reports that cap. Exact request_ids take precedence over either scope. This launch surface is deliberately self-scoped: it never exposes another credential's activity, prompts, SQL, financial results, access-token IDs, or warehouse job identifiers. Each successful lookup is itself a settled operation-based FIQ event even when it runs entirely in the control plane with zero warehouse billed bytes. A completed attributable operation is durably recorded before its successful result is returned. This lookup cannot include its own event because that event commits after the lookup snapshot; it can appear on the next lookup. This evidence describes PipeLedger-controlled FIQ and data-warehouse usage only. billing_record.status="pending" with zero delivery attempts means the record is queued for delivery to the billing system, which normally completes within minutes; it is not a failed or unsettled charge. It does not claim the customer's host-model token cost, workflow outcome, or business value. Every result is restricted to the calling organization and credential. Privacy and billing: this call records a private PipeLedger usage event containing the organization-scoped credential, tool, status, timing, rows, and PipeLedger compute/warehouse units. That usage event does not store the host conversation, prompt, assistant messages, generated SQL, or returned financial rows. Query reconciliation uses structural metadata and cryptographic fingerprints rather than readable filter or scope values.

pl_whoami

ChatGPT
Inspect this connection's identity and access: organization, credential label and ID, connection_instance_id, the date access ends (expires_at), role, allowed tools, effective operator grants, clearance, dimension scope, row-filter columns, allowed marts and ledger completeness. Use for connection troubleshooting or before presenting totals as organization-wide. pl_schema already includes the same whoami object; reuse it when available. If ledger.complete is false, explain the returned restrictions and do not present results as the organization's complete ledger. credential.expires_at is when this connection's access ends unless an Owner or Admin extends it; null means no end date is set. It is not the hourly token renewal. When it is near, tell the user the date and that an Owner or Admin can extend the access. The connection instance ID distinguishes simultaneous connections; labels do not prove which host is calling. Use pl_profile for the stable account profile used by host connection settings. A missing allowed tool is not granted to this connection. PipeLedger tools never post or change ERP records or initiate payments, regardless of clearance. If asked for these actions, explain that the capability does not exist; higher permissions or clearance cannot enable it. The separately authorized manual Unit Register records unit movements, not General Ledger entries, but those movements can affect unit-based ratios and derived metrics. Free identity and access diagnostics: no FIQ charge or billing preflight. Authentication, permissions, rate limits and private security auditing still apply. No bearer tokens or key material are returned.

Capabilities

Writes

Example Prompts

Click any prompt to copy it.

Similar Apps

Listed on MCP App Store

Listed on MCP App Store

Add this badge to your site or README. It links to this page.

HTML

<a href="https://mcp-app-store.keenethics-labs.com/app/pipeledger"><img src="https://mcp-app-store.keenethics-labs.com/badges/listed-on-mcp-app-store.svg" alt="Listed on MCP App Store" width="220" height="44"></a>

Markdown

[![Listed on MCP App Store](https://mcp-app-store.keenethics-labs.com/badges/listed-on-mcp-app-store.svg)](https://mcp-app-store.keenethics-labs.com/app/pipeledger)