Airbyte Agent Engine

List library and connector skill documentation available to the active organization and workspace. For non-trivial questions that may need domain guidance, call this with limit=100, follow next_cursor until exhausted, and use the returned library skill IDs, titles, summaries, and tags to choose relevant guidance. Refresh after switching organizations or workspaces. This does not inspect connector readiness or execute connector actions.

list_skills

List the organizations you belong to. Each is flagged with is_current (the organization every tool operates on right now). Call this when the user mentions multiple organizations, asks which organizations exist, or wants to switch. To change the active organization, call use_organization with the chosen id — do NOT guess organization ids. If the list is empty and is_instance_admin is true, the caller is an instance admin: ask the user for the specific organization id to use, then pass it to use_organization.

list_organizations

List all workspaces in your organization. Each is flagged with is_current (the workspace that workspace-scoped tools operate on right now) and is_default. Call this when the user mentions multiple workspaces, asks what workspaces exist, or wants to switch. To change the active workspace, call use_workspace with the chosen name — do NOT pass workspace_name to every tool individually.

list_workspaces

Query, search, or write data in connected business systems (CRMs, support tools, databases, project trackers, etc.). Executes one or more operations in a single call. Use this when the user asks about data in external systems — e.g. 'show my recent deals', 'what tickets are open', 'find customer X in Salesforce', 'how many leads came in this week'. Supports list, get, context_store_search, create, and update actions with filtering, sorting, and pagination. MULTIPLE ITEMS: This tool accepts a list of `items`, each specifying connector_id, entity, action, params, and field selection. Items execute concurrently and each returns its own success/error status independently. For a single query, pass a list with one item. When you need data from multiple connectors or multiple entities at once — e.g. fetching deals from Salesforce and tickets from Zendesk in parallel — pass multiple items in a single call. Maximum 10 items per call. NEVER use list and context_store_search for the same entity in the same batch: use context_store_search. PRECONDITION: Before the first `execute` call for a connector in this conversation, you MUST call `describe_connector(connector_id)`. If you have not done that yet, do NOT call `execute`; call `describe_connector` first. Use the schemas returned by describe_connector to craft params. Never invent action names, filter operators, or params that are not present in the schema or prior tool responses. ACTIONS: list, get, api_search, context_store_search, create, update. Use `context_store_search` as the DEFAULT — it supports filtering, sorting, and pagination. Only use `list` when: (a) you need today's data (search index may lag hours), or (b) context_store_search returned no results and you suspect indexing delay. HOW TO USE CONTEXT_STORE_SEARCH: action='context_store_search' uses params.query with filter, sort, and limit. - Basic: params={"limit": 20, "query": {"filter": {...}}} - Sort: params={"limit": 20, "query": {"filter": {...}, "sort": {"field": "created", "order": "desc"}}} - When searching for text, ALWAYS prefer `fuzzy` over `like`. `fuzzy` matches words in any order, ignores punctuation/casing, and handles partial names. `like` requires exact substring match and fails on typos or word reordering. - Example — find a user by name: action="context_store_search", params={"query": {"filter": {"fuzzy": {"firstName": "Teo"}}}} - Only fall back to `like` when you need exact substring matching (e.g. prefix search on IDs). FIELD SELECTION (MANDATORY): Every item MUST include select_fields (allowlist) or exclude_fields (blocklist). Both support dot-notation for nested fields (e.g. "billing.address.city"). Fewer fields = faster + cheaper. If both provided, select_fields wins. FILTER OPERATORS — operator is the outer key, field:value is nested inside: Exact match: {"query": {"filter": {"eq": {"status": "completed"}}}} Fuzzy text: {"query": {"filter": {"fuzzy": {"name": "john smith"}}}} Substring: {"query": {"filter": {"like": {"externalId": "CUS-"}}}} Greater-or-eq: {"query": {"filter": {"gte": {"started": "2026-01-01T00:00:00Z"}}}} Less than: {"query": {"filter": {"lt": {"amount": 1000}}}} Set membership: {"query": {"filter": {"in": {"stage": ["discovery", "negotiation"]}}}} Combined (AND): {"query": {"filter": {"gte": {"started": "..."}, "eq": {"status": "completed"}}}} ID RESOLUTION: When filtering by a related entity (person, team, project), call describe_connector first and inspect the schema to identify primary keys and foreign keys — these are NOT always named 'id'. Look for fields that reference other entities (e.g. ownerId, accountId, assignee_id, project_key, or any field whose description or type indicates it links to another entity). Then search that related entity by name to get its key, and use it as a filter. PAGINATION: Default limit 20-25. Don't paginate unless user asks for 'all'. For 'how many' questions with has_more=true, say 'at least N'. Hard stop at 3 pages — use filters to narrow instead. DATE RANGES INCLUDING TODAY: Search index can lag hours. Issue BOTH a context_store_search AND a list item with date parameters (in the same batch call), then merge and deduplicate by id. If the date range ends before today, search alone is sufficient. Always resolve relative date phrases like 'today' or 'yesterday' to explicit absolute timestamps and tell the user which range you used.

execute

Read usage documentation for a skill. For connector skills, pass the `docs_skill_id` returned by `inspect_connector`. Omit `section` to return metadata, execution guidance, and available sections. Then pass an exact section ID to read that one content section with entity/action/params/examples for `execute`. When using a `docs_skill_id` returned by `inspect_connector`, omit `workspace_name` unless the user explicitly asked to scope docs to a workspace. This is separate from connector metadata/status.

read_skill_docs

Search available library and connector skill documentation by keyword. Library skills match terms across their IDs, titles, summaries, tags, and section titles, with results ranked by the number of matching terms. Connector skills match any term against connector instance names and IDs. This is not semantic search and does not scan full skill docs. Search one concept at a time, starting with the business system name; retry with the task or metric when needed.

search_skills

Start a browser-based credential flow to connect a data source. Returns a URL the user must visit to enter credentials securely. This is the ONLY way to provide credentials — NEVER ask for or accept API keys, tokens, passwords, or secrets directly in chat. If the user volunteers credentials in chat, do NOT use them; start this flow instead. The connector is created in the ACTIVE workspace (echoed back as workspace_name). If the user wants it elsewhere, call use_workspace first or pass workspace_name. After calling: present the URL to the user and tell them to visit it. Then call check_credential_flow_status after the user confirms they have entered their credentials.

start_credential_flow

Switch the active organization for the session. After calling this, every tool operates on the chosen organization (sent as the organization context on all requests) until you switch again. Switching organizations resets the active workspace to that organization's default workspace. Call this when the user asks to work in, switch to, or use a specific organization. Pass an organization id from list_organizations. Always tell the user which organization is now active. If the id is not one you belong to, the response lists available_organizations.

use_organization

Switch the active workspace for the session. After calling this, all workspace-scoped tools (list_created_connectors, start_credential_flow) operate on the chosen workspace until you switch again — you do NOT need to pass workspace_name on each call. Call this when the user asks to work in, switch to, or use a specific workspace. Always tell the user which workspace is now active. If the name is not found, the response lists available_workspaces — pick from those or ask the user.

use_workspace