Current
MCP server
Connect AI agents to Doxa with OAuth, scoped tool access, and control over which projects they can use
Overview
The Doxa MCP server lets an AI agent work on your documentation sites through the Model Context Protocol. Clients such as Claude, Cursor, and ChatGPT connect to the server, sign in as you, and run docs workflows only inside the projects you select.
The server is remote. You do not install or run anything locally. The connection snippet on the new-project screen gives your client the server address.
The server exposes 34 tools. Each tool belongs to a scope, and the scopes you grant decide what the agent can do.
What a connected agent can do
| Area | What the tools do | Reference |
|---|---|---|
| Account and projects | Read your account, list and read projects, list GitHub repositories, create and delete projects | Account and projects |
| Docs pages | Read the docs structure, read and search pages, write pages, edit docs, upload images | Docs pages |
| Branches | List, create, inspect, diff, sync, preview, and delete branches | Branches |
| Site settings | Read and update site settings on a branch | Site settings |
| Publish and history | Publish a branch, list publications, restore a publication | Publish and history |
| Approval requests | Read, list, and cancel approval requests for consequential actions | Approval requests |
| Domains | List, add, check, and remove custom domains | Domains |
| Generation | Start, read, and cancel documentation generations | Generation |
Sign-in and access
Signing in uses OAuth. You approve the connection on the Doxa consent page, which opens in your browser.
If you are not signed in to the dashboard, sign in first. The consent page uses your dashboard session.
On the consent page, choose the scopes to grant and the projects the agent can use: all projects, or a list you pick.
A pending sign-in request is valid for 10 minutes. Each consent is single use.
Access tokens last 1 hour.
Refresh tokens last 30 days, or 14 days without use.
You can review and revoke any grant under Settings, in the Connected agents section.
For what each scope allows, see MCP Scopes and Approval Requests.
Team projects
Team projects work with the MCP server when you have access to them through your team role.
list_projectsincludes team projects. Each one includes itsteamname.get_accountreturns your teams and your role in each one.Your role limits what the agent can do. If your role does not allow an action, the tool returns
not_found. A viewer's agent can read a team project, but its writes return "not found".
For team roles and invitations, see Teams, roles and invites.
Limits to know
Note:
Writes to main are rejected through MCP. The agent must create a named branch, make its changes there, and publish the branch from the dashboard or through the publish_branch tool.
Consequential actions can create an approval request instead of running at once. Review pending requests in the dashboard or with the approval request tools.
Publishing and restoring a publication need the
publishscope.Starting a generation needs the
generation:runscope.
Connect an agent
Open the connection snippet
Go to the new-project screen in the dashboard and copy the snippet from the MCP connect card.
Add the server to your client
Paste the snippet into Claude, Cursor, or ChatGPT, following the steps for that client.
Complete OAuth sign-in
Start the connection from your client and sign in when the consent page opens.
Choose scopes and projects
On the consent page, select the scopes and the projects the agent can use, then approve.
Verify the connection
Ask the agent to list your projects. It should call list_projects and return the projects you selected.
For client-specific steps, see Connect an AI Agent with MCP.