ask-docs

Ask CrewAI Docs

Answer CrewAI questions from the official documentation, matched to the crewai version the user actually runs.

Verified against crewai 1.15.23 and the docs sites below on 2026-10-01. Live-tested with real LLMs on 2026-10-01.

There are two documentation sites:

Old docs.crewai.com/en/enterprise/... links redirect to docs-platform.crewai.com/platform/en/... - to the matching page when it still exists (as HTML, even from a .md URL), otherwise to the platform introduction. Look platform topics up on docs-platform.crewai.com directly; its pages also serve Markdown with a .md suffix.

When to Use This Skill

• A CrewAI feature, parameter, or behavior that the other skills do not cover
• Another skill says "re-verify with ask-docs" because the user's crewai version differs from the version that skill was checked against
• Current API syntax, method signatures, or configuration options
• An error message that may be covered by a troubleshooting note or caveat in the docs
• Less common features: telemetry, observability integrations, CLI commands, the tools library, AMP platform features
• Experimental conversational Flows (handle_turn(), ConversationConfig, RouterConfig, tracing, streaming)

Use a sibling skill first when it covers the topic. They hold curated guidance checked against crewai 1.15.x:

Use ask-docs for the gaps, and to re-check a sibling's version-sensitive rows.

How to Query the Docs

Use the first path that is available.

Path A: the docs MCP server (if configured)

The docs.crewai.com/mcp server needs no authentication. It exposes:

In Claude Code the tools appear as mcp__<server-name>__search_crew_ai (for example mcp__crewai-docs__search_crew_ai). The platform server has the same shape with a _platform suffix: search_crew_ai_platform, query_docs_filesystem_crew_ai_platform.

Rules:

• Always pass `version` with a leading v, e.g. "version": "v1.15.23". Without it, results mix edge and every hosted release (1.10.0 onward), often old ones first. "1.15.23" without the v returns "No results found".
• The filesystem is laid out as /<version>/<lang>/<path>.mdx (/v1.15.23/en/concepts/knowledge.mdx, /edge/en/...) plus /openapi/<version>/enterprise-api.en.yaml. rg -l "restore_from_state_id" /v1.15.23/en finds every page that mentions a symbol.
• Search results can be large (60 KB+). Prefer a focused query, then head -120 <file> on the page you need.
• A call occasionally fails with "Search failed". Retry once before falling back to Path B.

Path B: llms.txt and Markdown pages (no setup)

1. Fetch the index:

   WebFetch: https://docs.crewai.com/llms.txt

It is about 220 lines of - [Title](https://docs.crewai.com/v1.15.23/en/<path>.md): description, under ## Docs, followed by ## OpenAPI Specs (the AMP REST API YAML). The index has no category headings: find the page by title, or by path prefix (/en/concepts/, /en/guides/, /en/learn/, /en/tools/, /en/mcp/, /en/observability/, /en/api-reference/).

1. Fetch the page with the .md suffix, which returns clean Markdown (about 30 KB) instead of the rendered HTML (about 850 KB):

   WebFetch: https://docs.crewai.com/en/<path>.md

An unversioned /en/... URL redirects to the latest release (/v1.15.23/en/... on 2026-10-01).

1. For AMP platform questions, use https://docs-platform.crewai.com/llms.txt the same way. It is organised by heading (Getting Started, Build, Operate, Manage, Integration Docs, Triggers, How-To Guides, API Reference).

https://docs.crewai.com/llms-full.txt is every page in one file (about 2 MB). Do not load it into context. Download it and search it (curl -s ... | grep -n -A20 "restore_from_state_id") when the index does not point to the right page.

For conversational Flows, go straight to https://docs.crewai.com/en/guides/flows/conversational-flows.md. Treat it as the source of truth for the conversational API (imported from crewai.flow; crewai.experimental.conversational is a deprecated alias of the same module), which may still change.

Path C: GitHub source of the docs

The docs live in the public crewAIInc/crewAI repo under docs/<version>/<lang>/<path>.mdx, e.g. https://raw.githubusercontent.com/crewAIInc/crewAI/main/docs/v1.15.23/en/concepts/knowledge.mdx or .../docs/edge/en/.... The older docs/en/... layout no longer exists (404). Use this only when the docs site is unreachable.

Match the docs to the user's version

1. Run crewai version (or uv run python -c "import crewai; print(crewai.__version__)" in a uv project).
2. Read the docs for that release: the URL prefix https://docs.crewai.com/v<X.Y.Z>/en/<path>.md, or version: "v<X.Y.Z>" on the MCP server. Releases from 1.10.0 onward are hosted, plus edge (unreleased main). Only the root llms.txt exists - /v<X.Y.Z>/llms.txt is a 404 - so take the path from the root index and swap the version prefix.
3. If the user's release is not hosted, read the nearest hosted one and say so. A non-hosted version does not 404: /v1.9.0/en/concepts/knowledge.md redirects to the home page with status 200, so check that the page you got is the page you asked for.

Check docs snippets against the installed package

The docs are the best index of what exists, not proof that a snippet runs. For example, check decorator call forms such as @persist() against the installed version.

Before giving the user code that matters:

• Check signatures against the installed package: python -c "import inspect, crewai.flow.flow as m; print(inspect.signature(m.Flow.kickoff))".
• Run the smallest snippet that exercises the claim, if it needs no API key.
• If the docs and the package disagree, say so, give the version that works, and cite both. The check-crewai-api skill lists known differences between remembered and current APIs.

Workflow Summary

1. Understand the question - which concept, API, or behavior, and which crewai version?
2. Pick the site - framework (docs.crewai.com) or AMP platform (docs-platform.crewai.com).
3. Query - MCP with version set, or llms.txt then the .md page for that version.
4. Check it - signatures or a small run against the installed package when code is involved.
5. Answer and cite - give the docs URL (with the version prefix) so the user can read further.

Worked examples (checked 2026-10-01 against crewai 1.15.23)

Other good uses:

Setting Up the Docs MCP Server (optional)

Path B works with no setup. For structured search, add the server to the coding agent.

Claude Code (claude mcp add):

claude mcp add --transport http crewai-docs https://docs.crewai.com/mcp
# --scope project writes .mcp.json in the repo instead; --scope user makes it global
claude mcp list   # shows "crewai-docs: https://docs.crewai.com/mcp (HTTP) - Connected"

The project-scope .mcp.json it writes:

{
  "mcpServers": {
    "crewai-docs": {
      "type": "http",
      "url": "https://docs.crewai.com/mcp"
    }
  }
}

Codex CLI - codex mcp add crewai-docs --url https://docs.crewai.com/mcp, which writes this to ~/.codex/config.toml (or $CODEX_HOME/config.toml):

[mcp_servers.crewai-docs]
url = "https://docs.crewai.com/mcp"

Cursor - add to .cursor/mcp.json in the project, or ~/.cursor/mcp.json for all projects (Cursor MCP docs):

{
  "mcpServers": {
    "crewai-docs": { "url": "https://docs.crewai.com/mcp" }
  }
}

Other agents: add https://docs.crewai.com/mcp as a remote (streamable HTTP) MCP server. Add https://docs-platform.crewai.com/mcp the same way for AMP platform docs.

Related Skills

• getting-started — project scaffolding, choosing abstractions, Flow architecture
• design-agent - agent Role-Goal-Backstory, parameter tuning, tools, memory and knowledge
• design-task — task descriptions, expected_output, guardrails, structured output, dependencies
• check-crewai-api - current 1.15.x API versus the 0.x API assistants remember
• build-flow - Flow state, decorators, routing, persistence
• connect-tools-and-mcp - custom tools, crewai_tools, MCP servers for agents
• test-crewai-project - deterministic pytest testing of crews and flows
• deploy-to-amp - getting a crew or flow onto CrewAI AMP
• call-deployed-crew - calling a deployed crew over HTTP