One endpoint per site, https://mcp.citeworthy.io/mcp, on any Pro or Studio plan. Read tools are always on; five reversible write tools open up with a write-scope credential. This page is the whole contract — what you can connect, what the agent can do, and what it deliberately cannot.
Settings → Connectors → Add custom connector, paste https://mcp.citeworthy.io/mcp, and authorize. Claude follows the discovery endpoints below on its own and opens our consent page, where you pick the site and the scope — no token to copy anywhere.
claude mcp add --transport http citeworthy https://mcp.citeworthy.io/mcp then approve the connection when Claude Code opens the consent page. Same flow as Claude.ai, from a terminal.
Mint a token in Settings → Connections (Pro+), then: claude mcp add --transport http citeworthy https://mcp.citeworthy.io/mcp --header "Authorization: Bearer cwk_…" — paste your own token in place of cwk_…; that prefix is a placeholder, not something to type literally.
Cursor speaks Streamable-HTTP MCP the same way: add a server pointed at https://mcp.citeworthy.io/mcp with either an Authorization: Bearer header carrying your token, or by letting Cursor's own OAuth support walk the discovery flow below.
A read-scope credential can list and read findings, proposals, applied overrides and the facts confirmed about your site — nothing that changes state. Add write scope and exactly five reversible tools open up:
Mute a finding, or lift the mute. Fully reversible by the other call.
Write a manual draft. Nothing reaches the live site until a human approves it in the dashboard.
Reject a draft. Publishes nothing; the page stays proposable on future runs.
Remove an applied override — the un-publish direction. Takes a value off the live site.
Four tools the engine defines are deliberately not granted here, each for a stated reason rather than an oversight:
Not exposed. Running an audit here would skip the plan's page cap, the in-flight guard and the DNS/SSRF check the dashboard's own "run now" applies.
Not exposed. It spends real third-party model quota against a pooled monthly allowance this endpoint does not yet consult.
Not exposed, and worth stating plainly: approving is publishing. The dashboard runs four extra guards (a robots.txt invariant, a structured-data re-check, a length re-check, the claim guard) that this endpoint does not have. An agent can revert a bad live value; it cannot be the thing that decides to go live.
Not exposed. It spends model quota on a Workers AI inference, and that binding is not present on this Worker at all.
Every tool call an agent makes through this endpoint counts against a monthly allowance: 10,000 calls a month on Pro and 50,000 on Studio. The allowance belongs to your account, not to a site — every site and every token you have spends from the same pool, so a busy site borrows headroom from a quiet one instead of running out on its own.
It is a hard stop, not an overage. At the cap a tool call comes back as an ordinary tool error your agent can read — a normal MCP tool result marked isError, carrying the number used, the allowance and when it resets, in plain English. That shape is deliberate: a transport-level failure would be surfaced by most clients as a connection fault with nothing actionable in it, and several would quietly retry it. Nothing is charged and nothing you asked for is half-done: the call is refused before it runs. Connecting still works, and so does listing the tools; it is the calls themselves that stop. The meter resets on the 1st of next month (UTC), and Settings → General shows what your account has spent this month and how much of it was this site.
If your team runs a Cloudflare One MCP Portal, add Citeworthy the same way you'd add any other server: paste https://mcp.citeworthy.io/mcp as the endpoint. OAuth is handled by the discovery documents below — the portal reads them, registers itself, and sends whoever connects to our own consent page to pick a site and a scope. There is nothing else to configure on our end, and nothing of ours to install on Cloudflare's.
GET https://mcp.citeworthy.io/.well-known/oauth-authorization-server
GET https://mcp.citeworthy.io/.well-known/oauth-protected-resource
Standard RFC 8414 / RFC 9728 documents; dynamic client registration
(RFC 7591) is open at POST https://mcp.citeworthy.io/oauth/register —
no manual app approval on our side.
We are not submitting this server to a public MCP registry or directory today — that's a separate piece of work. Everything above (the URL, the discovery documents) is enough for a client to connect without one.