MCP server for your docs
Where is it?#
Each project serves its own endpoint:
https://docsbook.io/<owner>/<repo>/api/mcp/serverWith a custom domain connected, the same server also answers at https://<your-domain>/api/mcp/server. The Public MCP server card in Settings ▸ Domain & API shows your exact address and a Claude Code command to copy.
What can a reader's agent do with it?#
The server and its main tools carry your project's name, so an agent knows whose docs it is reading before its first call. The name is your repository's, in lowercase, with each run of other characters turned into one _: acme/api-docs serves a server called api_docs-docs.
| Tool | What it does |
|---|---|
search_<slug>_docs |
Finds the pages that answer a question, by meaning — or by exact words while search by meaning is off |
read_<slug>_doc |
Reads one page in full, so the agent quotes it instead of inferring it from a title |
get_<slug>_doc_outline |
Lists every page, for "what do these docs cover" |
docsbook_agent_activity |
Shows what the Docsbook agent recently did on these docs: pages read and written, translations run |
find_skill |
Returns the skills you placed on this door, then Docsbook's public skills catalog |
get_info |
Says which product the server is for and how to connect it |
find_widget, list_content_widgets |
Describe the widgets a Docsbook page or chat can show |
search_docsbook_docs, read_docsbook_doc, list_docsbook_docs |
Read Docsbook's own manual |
Agents that learned the older names search, read_doc and get_doc_outline still reach the same tools.
Connect it#
Replace the address with your own; the server name is the label the client shows.
claude mcp add --transport http api_docs-docs https://docsbook.io/acme/api-docs/api/mcp/server{
"mcpServers": {
"api_docs-docs": {
"url": "https://docsbook.io/acme/api-docs/api/mcp/server"
}
}
}{
"servers": {
"api_docs-docs": {
"type": "http",
"url": "https://docsbook.io/acme/api-docs/api/mcp/server"
}
}
}[mcp_servers.api_docs-docs]
url = "https://docsbook.io/acme/api-docs/api/mcp/server"Cursor reads ~/.cursor/mcp.json or .cursor/mcp.json, VS Code reads .vscode/mcp.json, and Codex reads ~/.codex/config.toml. The agent then has your docs as tools, with nothing to sign in to.
What readers see on your pages#
The Copy page menu on every page carries two ways in, both on by default:
- Connect MCP — copies a prompt that installs your server in whatever agent the reader uses, and tells the agent to search first and read the page before relying on it.
- Connect to VSCode — installs the server in VS Code with one click.
Hide either one with the toggles on the Copy page menu card in Customize ▸ Content.
This is your readers' server. Your own connection is https://docsbook.io/api/mcp/server: signed in, it reaches your projects' settings and hands work to the Docsbook agent — see Tell your agent, get discovered.
What stays private#
The server reads your published pages, plus any skills you choose to place on it:
- Nothing writes. No tool changes a page, a setting or a source.
- Nothing behind the site is read — not your analytics, settings, sources or the agent's memory.
- A private site stays private. The endpoint still answers, but every tool that reads your pages refuses with
PRIVATE_DOCS.
FAQ#
Does a reader's agent cost me anything? Calls are not billed one by one. With search by meaning on, each question is embedded on your balance; if that balance is empty, search_<slug>_docs answers INSUFFICIENT_BALANCE while reading pages keeps working.
Can I switch the server off? There is no switch for the endpoint itself — it serves the same pages your public site does. You can hide Connect MCP and Connect to VSCode, or make the site private.
How is this different from llms.txt? llms.txt is an index an AI reads to find your pages; the MCP server lets an agent search and read them while it works. See llms.txt and Markdown for AI.