Give Claude, Cursor and ChatGPT access to your team wiki (BookStack MCP)

Why connect agents to a team wiki

A coding agent that cannot see your runbooks guesses. A chat assistant that cannot see your procedures answers from the public internet. Most teams already keep the answers in a wiki; the problem is getting them to the agent without copying pages into every prompt.

The useful jobs are ordinary ones:

• Look up the deployment checklist before changing a server.
• Answer a question from the handbook, with the page it came from.
• Draft or update a page after a change, for a human to review.
• Find pages that contradict each other or are out of date.

The same wiki then serves people and agents. We call that a human-agent collaboration wiki: agents read and suggest, people decide.

What MCP is

The Model Context Protocol (MCP) is an open protocol that lets an AI application call tools on an external server. The server describes its tools, such as “search” or “read page”, and the client lets the model use them. A remote MCP server is reached over HTTP (the “Streamable HTTP” transport), so nothing has to be installed on the agent’s machine except the client itself.

MCP does not decide what an agent may see. That is up to the server and the credentials the client sends. For a wiki, the sensible answer is to reuse the permissions the wiki already has.

What BookHost offers (beta)

Every BookHost workspace has its own MCP server at https://<workspace>.bookhost.co/mcp. Agents authenticate with a BookStack API token sent as Authorization: Bearer <token id>:<token secret>. In the dashboard, owners and admins create a dedicated agent user with a BookStack role of their choice; the token is shown once and can be revoked at any time.

The server offers read tools (search, shelves, books, pages, revisions, attachments) and, depending on the write mode, tools to create, update and append to pages or send a change for review. The full list and the configuration for each client are on the agent access page.

Which clients work today

To say it plainly: despite the title of this guide, ChatGPT and the claude.ai website cannot connect to a BookHost wiki today. OAuth sign-in is planned, without a committed release date. If you use OpenAI or Anthropic models, the Codex CLI and Claude Code or Claude Desktop are the working routes for now.

The security model

• The agent is a BookStack user. Every call runs as the user that owns the token, with that user’s role and permissions. Admin and guest roles cannot be given to agents created in the dashboard.
• A token works on one workspace. The workspace is chosen by the address, never by the token.
• Limits and switches. 120 requests per minute, 1,500 per hour and 60 writes per hour per token. Owners and admins can revoke one token or switch agent access off for the whole workspace.
• An activity log without content. It records the tool, the page id, the result and the latency, never page content, search terms or tokens.
• Page content is untrusted input. Anyone who can edit a page can write text aimed at an AI. The server labels page content as data, not instructions, but no server can guarantee how a model reacts. Be careful with agents that can also act on other systems.
• No AI processing by BookHost. BookHost does not send page content to an AI model for this feature. It goes to the client you connect, and from there to whichever AI provider that client uses, which you choose and are responsible for.

Reviewing agent edits: “Propose only”

Each workspace has one write mode: Off (read only), Propose only (the default) or Direct. In Propose only, an agent that creates or updates a page does not touch BookStack. Its change arrives in the same review queue as uploaded documents, marked as an agent proposal. An owner or admin reads it, edits it if needed, and approves or rejects it.

Direct mode suits agents you trust for a narrow job, such as keeping a generated reference page current. Their edits appear in the page’s revision history under the agent’s own name, so you can see and roll back what they did. An agent can still send an individual change to review with propose_change.

The self-hosted alternative

If you run BookStack yourself, you do not need BookHost for this. BookStack has a documented REST API with token authentication, and there are community MCP servers for BookStack that you can run next to your instance, or you can write a small one yourself. Check any third-party server’s code before giving it a token.

With a self-run server, the following is your job:

• Hosting the MCP server over HTTPS, or running it locally per user.
• Creating a limited BookStack role and user for the agent.
• Storing and rotating tokens.
• Rate limiting and logging without storing page content.
• Deciding how agent edits are reviewed before they go live.

The trade-offs of running BookStack itself are in our self-hosted vs managed comparison.

llms.txt: a map for agents

Besides MCP, each workspace publishes an llms.txt file at https://<workspace>.bookhost.co/llms.txt. It tells an agent where the MCP endpoint is and how to authenticate. It lists books and pages only if the workspace allows logged-out access, and then only what a logged-out visitor can see.