Pinax

MCP tools

The five tools Pinax exposes, their argument shapes, and what they return.

A pinax serve instance exposes five MCP tools. Every tool except list_docs accepts an optional docs argument to scope the call to a single manifest when more than one is loaded. Omit docs to fan out across every manifest.

list_docs

{ "name": "list_docs", "arguments": {} }

Returns one entry per loaded manifest:

[
  { "name": "stripe", "url": "https://docs.stripe.com", "pages": 412 },
  { "name": "react", "url": "https://react.dev", "pages": 173 },
]

Use it first if you don’t already know what’s mounted.

list_sections

{
  "name": "list_sections",
  "arguments": { "docs": "stripe" },
}

Returns URL paths grouped by top-level section with page counts:

[
  { "section": "/api", "pages": 138 },
  { "section": "/payments", "pages": 67 },
  { "section": "/billing", "pages": 41 },
]

Cheap call - sections are derived from the manifest in memory.

search_pages

{
  "name": "search_pages",
  "arguments": {
    "query": "payment intent confirm",
    "limit": 10,
    "docs": "stripe",
  },
}

BM25 ranked over URL paths, titles, and section names, with a substring + fuzzy fallback so typos still surface results. Returns:

[
  {
    "url": "https://docs.stripe.com/api/payment_intents/confirm",
    "title": "Confirm a PaymentIntent",
    "section": "/api/payment_intents",
    "score": 0.84,
  },
]

Omit docs to search every loaded manifest in one shot. Results are interleaved by score, not grouped per docs site.

get_section_pages

{
  "name": "get_section_pages",
  "arguments": {
    "section": "/api/payment_intents",
    "docs": "stripe",
  },
}

Returns every page whose URL starts with that section prefix. Useful when the model wants to walk a topic exhaustively.

get_page

{
  "name": "get_page",
  "arguments": {
    "url": "https://docs.stripe.com/api/payment_intents/confirm",
  },
}

Live-fetches the URL with Accept: text/markdown first, falls back to HTML extraction, and returns clean Markdown. The page cache is consulted first - entries older than the TTL trigger a refetch.

get_page is the only tool that hits the network at call time. The others read from the in-memory manifest.

Error contract

Every tool returns a structured error rather than throwing:

{
  "error": {
    "code": "not_found",
    "message": "no manifest named 'striipe' (did you mean 'stripe'?)",
  },
}

Codes you’ll see: not_found, invalid_args, fetch_failed, cache_error, internal.

Edit this page on GitHub
Pinax