MCP server — your scan inside Claude, Cursor and your coding agent
The last step of the action plan used to be a to-do: change the robots rules, add the schema, write llms.txt, publish the page. Now the agent that already edits your site can do it. CiteGraph is a remote MCP server: Claude.ai and Claude Desktop connect as a custom connector by signing in, and Claude Code, Cursor, Windsurf, VS Code and Codex connect with a key. The agent reads your blind spots, the pages the engines cite, the fifteen readiness checks with their fixes and the pages written to win, then applies the changes in your repo.
How it works
Connect in a minute
In Claude: Settings, Customize, Connectors, Add custom connector, paste https://www.citegraph.app/mcp, sign in, allow. For a coding agent: create a key under API & MCP, paste the snippet for your client. Setup for every client is on the docs page.
Say what you want
"Fix my AI readiness" and the agent calls get_readiness, applies the failing checks' fixes with the code blocks as given, shows you the diff, and marks the moves done once shipped. "Why am I not named on the agency question" and it reads the blind spot and the verbatim answers.
Within your plan, never in public
Reads are unlimited within a fair-use cap. Drafts, pages and scans count against the plan exactly as in the app. It never touches billing and never posts anywhere public. Keys carry your role; a view-only member's key cannot change anything.
In depth
Twenty-three tools
Projects, overview with margins, blind spots with verbatim answers, answers per engine, cited sources, the action plan with pasteable fixes, readiness for a project or any domain, move status, drafts, pages (list, get, write, fill gaps), scans (run, status), memory (get, remember, forget) and the tracked questions (list, add, edit, remove).
Standards, not a bridge
Streamable HTTP, stateless, so it works from any machine. OAuth 2.1 with dynamic client registration and PKCE for clients that sign in; hashed API keys shown once for the rest. Any client that follows the MCP authorization spec connects the same way.
The same tools as Ask
The in-app assistant runs on this tool layer. A fix in one place serves both, and what your agent does shows up in the app at once.
What we won't do: run a scan behind your back. The agent can start one only when you ask, never forces a fresh one under a week old without you saying so, and it spends your plan's allowance, so it asks first.
Common questions
Which clients work?+
Claude.ai and Claude Desktop as a custom connector (sign in, no key). Claude Code, Cursor, Windsurf, VS Code Copilot agent mode and Codex CLI with an API key. Any client that speaks Streamable HTTP and either OAuth 2.1 or bearer keys.
Is it included in my plan?+
Yes, on every plan and the trial. 3,000 calls per key per day; thirty fresh readiness checks of arbitrary domains per key per day.
Is it safe to give an agent?+
Keys and tokens are hashed at rest and revocable from the API page; the agent stops at the next call. Keys carry your role. The server cannot change billing or post in public. Per-project scoping of keys is not there yet: a key sees the account's projects.
See it on your own domain — first dossier in about a minute.