Docs MCP — read the docs as tools

The Docs MCP exposes the same documentation the site renders — but as tools an agent can call instead of HTML to scrape. It is for the integration developer agent: the one writing a client against the API, which would rather search_docs("verify webhook signature") than fetch and parse a page.

It is public, read-only, and stateless — no token, no sandbox, no slot.

Why public?

The corpus it returns is already public, byte-for-byte — the same markdown the docs portal and /llms.txt serve. Gating a read-only copy of public content would only add friction, so it is ungated. (The Rails MCP, which mutates per-tenant state, stays token-gated — a deliberate public-read-only-docs vs. gated-stateful-runtime split.)

Connect it

https://partifact-docs-mcp.thanhvuttv.workers.dev/mcp

1. No token, no slot — just paste the URL. In claude.ai: Settings → Connectors → Add custom connector.

2. Once connected, sanity-check it: ask the agent to run search_docs("verify webhook signature"). It should return the Webhooks page.

Liveness (no token needed):

curl https://partifact-docs-mcp.thanhvuttv.workers.dev/

The 3 tools

Tool	Args	Returns
list_docs	—	Every page: { path, title, summary }. Read this first to plan which pages you need.
get_doc	path	One page as raw markdown (.md / docs/ prefix accepted). A miss returns an enriched error listing the available paths.
search_docs	query	Best-matching pages, ranked by term density, each with a snippet.

Each tool returns structuredContent plus a one-screen text summary (parity with the Rails MCP).

When to use it vs. the docs site

They serve the same corpus through two channels — pick by who is reading:

• A human, or an agent that can browse → the rendered docs portal, or /llms.txt / /llms-full.txt for a single-fetch dump.
• An agent mid-task that wants targeted retrieval → the Docs MCP. search_docs surfaces the right page without spending context on a crawl.

It's also an experiment

The Docs MCP is arm 3 of a small developer-experience experiment: does a coding agent build an integration against the API faster with a docs MCP than with the markdown portal + llms.txt alone, or with HTML-only docs? The corpus is held byte-identical across arms so the comparison is the access channel, not a privileged copy. See Let the AI do the coding.

Run it locally instead

The Docs MCP also runs over stdio (content is bundled — no backend). See Claude Desktop and Claude Code for the local config.