Read the docs over MCP

Connect an agent to Amdahl's documentation server so it can list, read, and search these pages without leaving its tool surface

Amdahl publishes its documentation as its own MCP server. An agent connected to it can list every page, read one in full, and search across all of them, without a browser and without being handed a URL.

The server URL is:

code

https://docs.amdahl.ai/mcp

It is public. There is no API key, no OAuth, and no workspace behind it — it serves exactly the pages you are reading now.

Why it is a second server

The workspace server at https://app.amdahl.ai/mcp carries your data: your conversation corpus, your evals, your connections. The docs server carries the manual. They are deliberately separate, for two reasons.

Auth. Your data needs a credential. The docs do not, and requiring one would mean nobody could read how Amdahl works until after they had signed up.

Tool budget. Every tool definition an agent holds costs context on every call. Keeping the docs off the workspace server means connecting for data costs you three tool definitions you did not ask for, and connecting for docs costs you nothing on the data side. Connect one, the other, or both.

What your agent gets

Three tools, all read-only.

Tool	What it does
`list_docs`	The full page tree — every slug, its title, and the section it sits in. Start here when you do not know what exists.
`read_doc`	One page as markdown, by slug. The same bytes the site serves.
`search_docs`	Scored full-text search across every page. Returns matching slugs with the surrounding excerpt.

The same pages are also exposed as MCP resources under doc://<slug>, so a client that prefers resources/list and resources/read can use those instead of the tools.

Connect it

Adding this alongside the workspace server is the intended setup. The two do not conflict — different origins, different tools, and only one of them needs a key.

Claude Code

bash

claude mcp add --transport http amdahl-docs https://docs.amdahl.ai/mcp -s user

Claude Desktop / Cursor — add to your MCP config:

json

{
  "mcpServers": {
    "amdahl-docs": {
      "type": "http",
      "url": "https://docs.amdahl.ai/mcp"
    }
  }
}

OpenAI Responses API — as a remote MCP tool:

json

{
  "type": "mcp",
  "server_label": "amdahl_docs",
  "server_url": "https://docs.amdahl.ai/mcp",
  "require_approval": "never"
}

No headers block: there is nothing to authenticate.

The Amdahl plugin already includes it. If you installed amdahl-gtm@amdahl-cookbook, you have both servers.

If your client cannot add a second server

Every page has a markdown twin at its own URL. Point the agent at https://docs.amdahl.ai and it will get markdown rather than HTML when it asks for it:

bash

curl -H "Accept: text/markdown" https://docs.amdahl.ai/mcp/docs-server

https://docs.amdahl.ai/llms.txt is the curated index of the whole site, which is the right thing to read first if you are fetching rather than connecting.

Freshness

The docs server reads the same published snapshot the website renders, so it is exactly as current as docs.amdahl.ai — never staler, never fresher. If a page reads correctly in a browser, the MCP server returns the same content.