Read these docs with an agent
A coding agent reads documentation badly through a browser and well through Markdown. Every page on this site, the help centre and the distributor site is published twice: as the page you are reading and as the Markdown it was rendered from. There is an index of each site for an agent to start from, and a small MCP server that serves the same pages as tools.
Who can use it: anyone who can reach the documentation. Nothing here needs a platform licence or an API key.
The Markdown twin of every page
Add .md to any documentation path and you get the source of that page.
| What you want | Where it is |
|---|---|
| One page as Markdown | The page's path with .md, for example /cookbooks/ask-a-knowledge-graph.md |
| An index of the site, with a description per page | /llms.txt |
| Every page of the site in one file | /llms-full.txt |
llms.txt is a list of links grouped by section, which is the cheap way for an agent to see what exists before fetching anything. llms-full.txt is the whole site in one response, which is the expensive way and is better suited to a one-off ingest than to a question.
The Markdown is the rendered page, so the generated reference pages appear in it with their tables intact, and a page's licence pills appear as a line of feature names under the title.
The docs MCP server
The same pages, as tools a coding agent can call rather than files it has to fetch. The server is in the documentation repository and runs beside your checkout over stdio.
pnpm build # the server reads the built Markdown
node tools/docs-mcp/server.mjs # speaks MCP on stdin and stdout
Point a client at it the way you point it at any stdio MCP server:
{
"mcpServers": {
"katonic-docs": {
"command": "node",
"args": ["tools/docs-mcp/server.mjs"],
"cwd": "/path/to/katonic-docs-platform"
}
}
}
| Tool | What it does |
|---|---|
search_docs | Finds the pages containing every word of a query, best first, each with a snippet and a path. Takes query, and optionally site and limit |
get_page | Returns one page as Markdown, exactly as published. Takes the path from a search result |
list_pages | Lists paths and titles, optionally filtered by site and by a path prefix |
A session usually goes search, then read: search_docs with the question, then get_page on the path that looks right. list_pages with a prefix such as /cookbooks/ is the quickest way to see one section whole.
What the search does and does not do
The search is lexical. It requires every meaningful word of the query to appear on the page, weighs a word in the title far above a word in the body, and damps long pages so that a generated reference page does not outrank the guide that answers the question. It does not embed anything, so a question phrased in words the documentation never uses will find nothing rather than finding something close.
When a search returns nothing, drop to the distinctive noun: rerank, agent_type, no_models_for_requested_tier. Those are the words the pages actually contain.
Limits
- The server reads the built Markdown, so it answers from the last
pnpm buildrather than from the working tree. Rebuild after editing a page. - It is a local server against a checkout. There is no hosted endpoint, and it holds no credentials and reads nothing but the built files.
- Search covers the pages of all three sites at once. Pass
sitewhen you want only one; a question about the public API can otherwise match a help centre page that uses the same words. - Stop words are dropped from a query, so a search for "how do I" alone matches nothing.
- The
.mdtwin andllms.txtcome from the site build, which means they exist wherever the site is served from, including a localpnpm serve.