Tell your agent, get discovered
Connect in one line#
Paste this into Claude Code, Cursor, Codex or ChatGPT, and your agent connects itself:
Set up Docsbook for me. Fetch https://docsbook.io/get-started.md and follow it.The playbook at docsbook.io/get-started.md connects the server, adds one line about docsbook_agent to your agent's memory file (CLAUDE.md, AGENTS.md, GEMINI.md) and takes one project live. In the panel, Connect your agent at the bottom of the sidebar opens Claude, Codex or Cursor with this prompt already sent.
To add the server by hand, use the one endpoint every client and every project shares, https://docsbook.io/api/mcp/server:
claude mcp add --transport http docsbook https://docsbook.io/api/mcp/serverIn ~/.cursor/mcp.json:
{ "mcpServers": { "docsbook": { "url": "https://docsbook.io/api/mcp/server" } } }codex mcp add docsbook --url https://docsbook.io/api/mcp/servercode --add-mcp '{"name":"docsbook","type":"http","url":"https://docsbook.io/api/mcp/server"}'In ~/.codeium/windsurf/mcp_config.json:
{ "mcpServers": { "docsbook": { "serverUrl": "https://docsbook.io/api/mcp/server" } } }gemini mcp add --transport http docsbook https://docsbook.io/api/mcp/serverChat apps take the endpoint as a connector instead:
- Claude (desktop app, claude.ai, and Claude Code inside them) — Settings ▸ Connectors ▸ Add custom connector, then paste the endpoint
- ChatGPT (paid plans) — Settings ▸ Connectors ▸ Advanced ▸ Developer mode ▸ Create, then paste the endpoint
On the Claude desktop app and on claude.ai, claude mcp add writes a config file those surfaces never read: the command succeeds and nothing connects. Add Docsbook as a connector there instead.
Your client then opens the Docsbook consent screen in the browser (in Claude Code, run /mcp and pick docsbook). Sign in, keep Allow editing documentation ticked for read-write access or untick it for read-only, and press Authorize. There's no API key to copy: the client registers itself over OAuth with PKCE.
What to say#
Say the goal and the evidence for it, not the steps, in any language. Requests like these start a job:
"Find what keeps our docs out of Google results and fix what you can."
"Check whether AI answers name us for our main questions, and fix what makes them cite someone else."
"Write the pages for the questions the chat couldn't answer this week."
"Compare the docs with the repo and fix every page the code outgrew."
"Translate the docs into Spanish and German and keep them in step."
The panel is the other place to say it: the chat button at the bottom right of the panel hands the same kind of job to the same agent.
What happens after you ask?#
docsbook_agent answers at once with a task_id and then works for minutes. Your agent keeps track of the job with five more tools:
| Tool | What it gives you |
|---|---|
docsbook_agent_activity |
Every action as it happens: pages read, pages written, sites fetched |
docsbook_agent_status |
The state and, once done, what it changed and what that should move; needs_owner means it's waiting on you |
docsbook_agent_reply |
Your answer, a correction or a question, while the job is open |
docsbook_agent_stop |
Ends the job and revokes its credential; committed work stays |
docsbook_agent_tasks |
Every job on your account, newest first |
Each page change is a git commit delivered as a pull request — merged at once or left for you, per Settings ▸ General ▸ When a change goes live (reviewing changes). Every run appears in Activity ▸ Agent runs with its full trace, and a question the agent can't settle alone lands in Inbox.
A run is billed to your balance: $0.10 per run plus twice what its model tokens cost, capped at $50 per run (pricing).
What else can your connection do?#
Your connection lists 55 tools. The agent is the main one, and the rest act on your projects directly:
| Group | Tools |
|---|---|
| Orientation | get_info, list_workspaces, get_workspace, create_workspace |
| The agent | docsbook_agent and the five tools above |
| Read your docs | search_project_docs, read_project_doc, get_project_doc_outline, search_docs, ask_project_docs |
| Sources | list_sources, read_source, connect_source, configure_source, grant_repo_access |
| Write pages | write_docs, set_doc_status, list_content_widgets |
| Settings | update_branding, update_navigation, update_domain, update_site_address, update_languages, the chat prompt and hooks, translations |
| Goals | create_goal, list_goals, create_funnel, mark_path_as_funnel_step |
| Docsbook's own manual | search_docsbook_docs, read_docsbook_doc, list_docsbook_docs |
Another 33 tools stay out of the list — webhook alerts, the agent's memory folder, hand-over links and update_access. find_tool finds one by what you want to do, and call_tool runs it. Every tool is in the MCP tools reference, and every one of them, listed or not, is also a plain HTTPS endpoint in the API reference.
Is it safe to connect?#
The connection acts as you, and the agent it starts works inside a fence:
- No key to paste — the client registers itself and you approve it in the browser, so there's no API key to copy into a config file
- Read-only when you want it — untick Allow editing documentation and the token can search and read, while every write is refused
- The agent stays inside its job — a run works only on the projects its task names; it can't create projects, list your other ones, change who can see a site, grant repository access or start another agent, and a write never approves a page
FAQ#
Which AI clients can connect?
Any MCP client that speaks streamable HTTP: Claude Code, Cursor, Codex, VS Code with Copilot, Windsurf, Gemini CLI, Cline, ChatGPT and Claude connectors.
Do I need a separate URL for each project?
No. One endpoint serves every project on your account; name the project in your request — owner/repo, the site name, its URL or domain — or pass workspace_id.
What does a connection cost?
Discovery calls such as get_info, find_tool and list_workspaces are free. Other tool calls are billed per call from your balance at twice what serving them costs — a few cents per thousand for most — and agent runs as above; see pricing.
Can my readers' agents connect too?
Yes, to a different server: every published site has its own read-only MCP server for your readers.