agentic-search-fragments
ChatGPTReturns data without cards; only explicit visual requests use view-memory-fragments. Primary discovery tool using vector-based semantic search (understands meaning, not keyword matching)—run it before starting work, when resuming, and any time context feels thin. Provide workspace scope (workspaceId or workspaceIds) and include repo:<name> tags when they help focus results (drop them if coverage is lacking). Use searchIntent, context, and related fields to spell out what you're trying to accomplish so the search can adapt (quick lookup vs. deep research). If the tool returns decision: NO_RESULTS, no relevant fragments were found—use the suggestedQueries array to retry with different search terms immediately. If the tool returns decision: NEED_MORE, adjust inputs or intent and rerun until it's satisfied before executing work, but stop once invocationLimitReached flips to true and document the uncovered gaps instead of looping. Reset the invocation attempt counter whenever you pivot to a substantively different search objective. Summarize selected fragments back to the user and always fetch full content with get-memory-fragment-content before relying on any returned fragment so you have complete context. Use collectionId to scope results to a specific collection. Use recencyWeight (0–1) to bias toward recently updated fragments for time-sensitive queries (0 = pure similarity, 1 = pure freshness). Important: Always phrase query and searchIntent as natural language—describe what you need as if talking to a knowledgeable colleague, not a keyword search engine.
attach-file-to-fragment
ChatGPTAttach a file to a memory fragment. The file must exist and be accessible. The fragment must exist and the user must have update permissions.
close-tabular-fragment
ChatGPTClose a dataset query session and release its execution resources. This does not delete the dataset or the original file.
create-collection
ChatGPTCreate a new collection in a workspace to organize memory fragments into logical groups. Collections have unique names within a workspace. Required permission: fragments.create.
create-fragment-symlink
ChatGPTCreate a read-only symlink of a fragment in another workspace. The child fragment will auto-sync when the parent is updated and will be cascade-deleted when the parent is deleted. Requires owner/collaborator in both source and target workspaces. Required permission: fragments.update. If the source type is absent in the target, workspace.manage_fragment_types is also required.
create-fragment-type
ChatGPTCreate a custom fragment type in a workspace. Changes are processed asynchronously (returns status 'processing'). Use get-fragment-types to list existing types first and avoid duplicates. Required permission: workspace.manage_fragment_types (owner role).
create-memory-fragment
ChatGPTDocument verified insights, implemented changes, and proven solutions as soon as they're confirmed. Include repository context, apply the Fragment Type Selection Guide, and surface residual risks in the content. YAML frontmatter support: Content can include YAML frontmatter (between --- delimiters) which will be automatically extracted and stored as searchable metadata. Frontmatter can be queried using JSONB operators in list-memory-fragments (e.g., frontmatter ? 'author' or frontmatter->>'version' = '1.0'). Required permission: fragments.create.
create-workspace
ChatGPTCreate a new workspace to organize memory fragments and collaborators. If you belong to an organization the workspace is created under it; otherwise it is a personal workspace. Creation is processed asynchronously (returns status 'processing'). Required permission: workspace.create (owner role).
get-collection
ChatGPTGet details of a specific collection by ID. Returns collection metadata including name, description, and timestamps. Optionally includes a paginated list of fragments in the collection when includeFragments is true. Required permission: workspace.read.
get-file
ChatGPTGet file metadata by ID. By default, includes a short-lived (5-minute) presigned download link. Security: the download link is a single-use secret — relay it directly to the requesting end-user, never log it, never include it in long-lived summaries, and never share it with other agents or sessions. Files are stored in Azure Blob Storage, so download links are used instead of returning file content to avoid filling the context window.
get-fragment-types
ChatGPTList active fragment types in the single required workspaceId, which must be accessible to the connected account. Returns each type's ID, name, description, icon, color, and default flags. Returns an empty list when that workspace has no active types; does not search other workspaces, create types, or validate a supplied fragment type ID. Required API Token Permission: workspace.read.
get-linked-fragments
ChatGPTRetrieve explicit deterministic linked_to edges on every serving plane. Hosted-zone Leaf additionally returns authorized projected symlink relations; inspect provenance.connectionType to distinguish them. Supports outgoing references, incoming backlinks, or both, with deduplicated results. Required API Token Permission: fragments.read.
get-memory-fragment-content
ChatGPTRetrieve data for reading, summarizing or analysis without displaying a card. Only explicit visual requests use view-memory-fragment. Retrieve fragment content by fragmentId (workspace access) or shareId (named recipient access to one fragment). A shared read returns title, content, and revisionToken only. Optional startLine/endLine or offset/limit enable partial reads. Required scope: fragments.read.
get-public-site-review
ChatGPTRead metadata, current eligibility and the private review URL for your own preparation. It does not return the public capability or publish content. Open the browser URL to inspect exact pages, navigation, images and metadata.
get-tabular-fragment-source
ChatGPTGet metadata and a short-lived download URL for the exact original file uploaded to a dataset fragment. The file is not converted and this operation does not run a query or consume query credits. The URL is a bearer secret for the requesting user only and must not be logged, persisted, summarized, or shared.
list-collections
ChatGPTList collections in a workspace with pagination. Returns collection metadata ordered by creation date (newest first). Required permission: workspace.read.
list-memory-fragments
ChatGPTPrimary data-only tool for listing, finding, and searching memory fragments. Ordinary listing does not request cards. Only explicit visual requests use view-memory-fragments. Use this tool when users ask to 'list fragments', 'find all fragments', 'show me fragments', 'get fragments', 'search fragments', or query fragments. Provides powerful SQL-like query capabilities with flexible date format support. Returns metadata only (title, summary, tags, dates, author, fragment type) without full content - use get-memory-fragment-content to fetch full content by ID. Capabilities: - SQL-like queries: Full SQL-like filtering with operators: =, !=, >, <, >=, <=, LIKE, ILIKE, IN, NOT IN, @> (JSON array/object contains), ? (key exists), ?& (all keys exist), ?| (any key exists), -> (get JSON field), ->> (get JSON field as text). All filtering is done through the query parameter. Example: status = 'active' AND fragment_type_id = 'uuid' AND tags @> ARRAY['nextjs'] AND created_at >= '2025-01-01' or frontmatter @> '{"author": "John"}' - Flexible date formats: Automatically converts multiple date formats to ISO 8601: ISO strings ('2024-01-15T00:00:00Z'), dates only ('2024-01-15'), timestamps (1705276800000), or human-readable ('2024-01-15 10:30:00', '2024/01/15'). Use date comparisons in queries: created_at >= '2025-01-01' or updated_at < '2025-11-01T00:00:00Z' - Operator precedence: Follows standard SQL precedence where AND has higher precedence than OR. Example: status = 'active' AND tags @> ARRAY['a'] OR tags @> ARRAY['b'] is interpreted as (status = 'active' AND tags @> ARRAY['a']) OR (tags @> ARRAY['b']). Tip: When in doubt about precedence in complex queries, use parentheses to explicitly group conditions. - Format guidelines: Use single quotes for string literals (e.g., 'active', '%test%'). For dates in queries, use any recognized format - they will be automatically normalized. - Complex queries: Supports nested conditions with parentheses, AND/OR operators, date ranges, tag filtering, text search (ILIKE for case-insensitive), status filtering, fragment type filtering, and combinations. Example: (status IN ('active', 'stale') OR status = 'archived') AND fragment_type_id = 'uuid' AND title ILIKE '%MCP%' AND created_at >= '2025-10-01' - Default behavior: If no query is provided, returns fragments with status 'active' or 'stale' (excludes archived). To include archived, use status IN ('active', 'stale', 'archived') in your query. - ORDER BY: Sort results by created_at, updated_at, or title with ASC/DESC. Example: created_at DESC or title ASC - Pagination: Use limit (1-100, default: 20) and offset (default: 0) for pagination Allowed query columns: id, fragment_type_id, title, summary, status, created_by, created_at, updated_at, tags, frontmatter (Note: workspace_id is NOT allowed - workspace filtering is enforced separately for security) Examples: - status IN ('active', 'stale') - Status filtering - fragment_type_id = '04a5fb62-1ba5-436c-acf7-f65f3a5ba6f6' - Filter by fragment type - tags @> ARRAY['repo:usable', 'mcp'] - Tag array contains (AND logic) - title ILIKE '%test%' - Case-insensitive text search - created_at >= '2025-10-01T00:00:00Z' AND updated_at < '2025-11-01T00:00:00Z' - Date ranges with timestamps - created_at >= '2025-10-01' - Date only (automatically converted) - created_at >= '1705276800000' - Timestamp format (automatically converted) - (status = 'active' AND title ILIKE '%MCP%') OR tags @> ARRAY['mcp'] - Complex nested queries - status = 'active' AND fragment_type_id = 'uuid' AND tags @> ARRAY['nextjs'] - Multiple conditions - frontmatter @> '{"author": "John"}' - JSON object contains - frontmatter ? 'author' - Key exists in frontmatter - frontmatter ?& ARRAY['author', 'version'] - All keys exist - frontmatter ?| ARRAY['author', 'version'] - Any key exists - frontmatter->>'author' = 'John' - Get field as text and compare - `frontma…
list-public-sites
ChatGPTList up to five sites within your current assigned publishing/site-management scope. Returns current release identifiers and versions, without secret links or private page bodies.
list-workspaces
ChatGPTList all accessible workspaces with role information, permissions, and metadata. Shows owned workspaces (full control) and invited workspaces (based on assigned role). Essential for understanding available knowledge contexts. Required API Token Permission: workspace.read.
manage-application
ChatGPTCreate, inspect, update, and manage OAuth credentials for marketplace applications. Use action to select the operation. Actions: - list — List all applications you own or admin - get — Get one application's full non-secret OAuth configuration (admin only) - get-mcp-profile — Get the app's MCP endpoint and shared-identity policy (admin only) - create — Register a new marketplace application with OAuth config - update — Update metadata, developer, branding, install access policy, token config, or OAuth configuration (admin only). Use changes.branding for invite branding (logoUrl, primaryColor, fromName, replyTo, loginUrl) and the account creation return origin (signupReturnOrigin); provided keys merge into the stored branding and null clears a key. Use changes.accessControl to change the install access policy (public, invite-only, approval-required, or workspace-restricted). Use changes.oauthUrlPatch for URL-only edits; it defaults to additive so production URLs are preserved, and updateMode: "replace" replaces each provided array. Use changes.oauthConfig to submit a full replacement config, which is required to change applicationType or grantTypes - update-permissions — Replace requested OAuth scopes using permissions, including fragments.delete with custom: false (admin only). Preserve existing scopes in the list. Active or beta apps require release.version and release.changelog. - update-mcp-profile — Fully replace the app's MCP endpoint and shared-identity policy using mcpProfile (admin only). Read the current profile first when preserving advanced overrides, then call get-mcp-profile after the accepted write to verify persisted values. Disabling blocks existing gateway calls; configuration changes require gateway owners to remove and re-add the app - reveal-credentials — Explicitly return the current client ID and secret without rotation (admin only; secret-bearing) - regenerate-credentials — Explicitly rotate and return credentials; the previous secret is immediately invalid (admin only; secret-bearing) - installation-spend-status — Inspect configuration prerequisites without secret material - emergency-revoke-installation-spend — Revoke every installation credential using app-admin authorization - list-webhooks, create-webhook (webhook.url, webhook.events; returns the signing secret once), delete-webhook, test-webhook (webhookId) — Manage app webhooks (admin only) - list-versions, publish-version (release.version, release.changelog, release.breakingChange) — Version history and releases (admin only) - list-admins, invite-admin (adminInvite), change-admin-role (targetUserId, adminRole), remove-admin (targetUserId), revoke-admin-invite, resend-admin-invite (inviteId) — Manage app admins and viewers (admin only) - list-installations (installationStatus: installed or pending), approve-installation, deny-installation (targetUserId) — Review install requests for approval-required apps (admin only) Required API Token Permission: fragments.create for ordinary application actions. Secret-bearing reveal-credentials and regenerate-credentials actions additionally require workspace.update. Installation-spend actions reject personal access tokens and require a user session plus application-admin authority in Core; emergency revoke is destructive.
manage-installation-spend-credential
ChatGPTInspect and control installation-bound background credit authority. Actions: status, list, approve, rotate, revoke. Core enforces installer or organization billing authority for each installation. approve/rotate create an app-redeemable one-time approval and return lifecycle metadata only. Responses never contain credential secrets, client secrets, or JWTs.
open-tabular-fragment
ChatGPTOpen an authorized, expiring query session for a ready dataset. Pass a returned sessionId to load another fragment from the same workspace. Returns every loaded fragment and collision-safe SQL table alias. Requires individual opt-in and an eligible subscription; data survives container sleep.
prepare-public-fragment
ChatGPTPrepare one Cloud Markdown/text fragment, selected safe images and public metadata for exact private review. Choose link-only or searchable access and a readable platform site address. Returns an authenticated human review URL; nothing is published. Assigned individual-source disclosure authority and publishing.prepare scope are required. This creates a source-bound one-page site within the five-site allowance; it does not grant arbitrary site or domain management. Inspect, request publication, list and withdraw through the public-site tools using the returned IDs. Ordinary fragment writes and tool confirmation booleans cannot publish.
prepare-public-site
ChatGPTPrepare a Cloud-hosted micro-site or a one-page fragment edition for exact private review. Nothing is published. Returns a private browser review URL; current assigned publishing or site-management authority and an explicitly granted publishing.prepare credential scope are required. Only Markdown/text and explicitly selected safe images are supported.
query-tabular-fragment
ChatGPTRun bounded read-only SQL over every fragment loaded into a dataset session. This consumes configured credits. Reuse requestId and exact arguments on retries. Returns at most 500 rows and 1 MiB; use aggregates or pagination for larger datasets. Access to every loaded fragment, opt-in and budget are checked for each operation.
remove-fragment-symlink
ChatGPTRemove a symlinked fragment by deleting the child copy. Requires owner/collaborator in the parent workspace. Required permission: fragments.delete.
request-fragment-import
ChatGPTRequest a short-lived, workspace-scoped import grant for hash-exact Markdown, HTML or tabular fragment creation (CSV, XLS, XLSX, XLSM, DuckDB, Parquet). Tabular files normalize asynchronously and require explicit opt-in and an eligible subscription. The document bypasses model context and is streamed unchanged to the returned HTTPS endpoint. The grant binds fragment metadata, byte size, SHA-256, user, workspace, and expiry. Required permission: fragments.create.
request-upload-url
ChatGPTGenerate a presigned upload URL for file uploads. Returns a URL that the user can upload to directly via curl/HTTP. The LLM should then poll with get-file(fileId) to check upload status.
search-files
ChatGPTSearch for files in a workspace by keyword, tags, or MIME type. Returns file metadata without file content.
search-memory-fragments
ChatGPTReturns data without cards; only explicit visual requests use view-memory-fragments. Fallback vector-based semantic search (understands meaning, not keyword matching). Use only when agentic-search-fragments errors or cannot reach a workspace. Provide the relevant workspace scope; add repo:<name> tags when they help narrow results and drop them if coverage is lacking. Review similarity scores and timestamps, then pivot back to the agentic tool for iterative expansion. Always fetch full fragment content with get-memory-fragment-content before relying on any returned fragment so you have complete context. Use recencyWeight (0–1) to bias toward recently updated fragments (0 = pure similarity, default). Required permission: fragments.read.
search-public-workspaces
ChatGPTSearch the public workspace registry — discover public workspaces you are not yet a member of or subscribed to, ranked by full-text relevance over name and description, with pagination. Use subscribe-to-workspace to subscribe to a result. Required API Token Permission: workspace.read.
subscribe-to-workspace
ChatGPTSubscribe to a public workspace to gain read-only access to its memory fragments. Use search-public-workspaces to discover workspace IDs. Subscription is processed asynchronously (returns status 'processing'). Only public workspaces you are not already a member of can be subscribed to. Required permission: workspace.subscribe.
update-collection
ChatGPTUpdate a collection's name or description. At least one field must be provided. Collection names must remain unique within the workspace. Required permission: fragments.update.
update-fragment-type
ChatGPTUpdate a custom fragment type's name, description, icon, or colour. System default fragment types cannot be modified. Changes are processed asynchronously (returns status 'processing'). Required permission: workspace.manage_fragment_types (owner role).
update-memory-fragment
ChatGPTUpdate a workspace fragment by fragmentId, or update only content on a named-user Editor share with shareId and the revisionToken from get-memory-fragment-content. Shared edits support content or patchOperations and cannot change metadata, type, or collections. Required scope: fragments.update.
update-workspace
ChatGPTUpdate an existing workspace's name, description, visibility, or trusted HTML runtime policy and exact HTTPS resource allowlists. Use htmlRuntime.resources.script/style/image/font/connect to configure allowed origins. Workspace-wide script execution requires confirmWorkspaceScripts=true. Changes are processed asynchronously (returns status 'processing'). Required permission: workspace.update (owner role).
view-memory-fragment
ChatGPTUse only when the user explicitly asks to view a fragment visually or display a fragment card. Ordinary read, fetch, summarize or analyze requests use get-memory-fragment-content. Reuses that tool's selectors, partial reads and access restrictions. Named-share reads disclose only the existing restricted shared result. Leaf supports local fragmentId or key plus workspaceId reads; shareId and partial reads require Cloud. Required scope: fragments.read.
view-memory-fragments
ChatGPTUse only when the user explicitly asks to display fragment cards or a visual table. Ordinary list, filter, count, search or analyze requests use list-memory-fragments or the search tools. Reuses list filtering, ordering and pagination with complete rows; select and groupBy are rejected. To display chosen search results on Cloud, filter by their exact returned IDs, e.g. query: id IN ('uuid-1', 'uuid-2'). Leaf requires workspaceId and supports unfiltered pagination only; query/orderBy require Cloud. Required scope: fragments.read.
withdraw-public-site
ChatGPTWithdraw the exact active site release, matching its identifier and version. Requires explicit publishing.withdraw credential scope and current assigned publisher authority. Withdraws every page and navigation entry; copied content elsewhere cannot be recalled.
