Orbilo now has an MCP server. Ask ChatGPT, Claude or Cursor whether a site is ready for agents.

Orbilo now ships a Model Context Protocol server. Point ChatGPT, Claude, Cursor, VS Code, Codex or Claude Code at https://orbilo.co/mcp and your assistant can check any site against UCP, WebMCP and Access, tell you what is failing and why, read the sites your team monitors, and explain a spec release from the actual changelog. Checking needs no account. Monitoring needs the one you already have, including on the Free plan.

The MCP page has the endpoint, a one-minute setup for each client and the full tool list. The docs go deeper. This post is about why we built it and what it is useful for.

Why an MCP server

The people who ask "can agents use our site?" are increasingly asking an AI assistant, not a search engine. The assistant does not know. It has training data about UCP from whenever it was trained, no idea whether your manifest is valid today, and no way to load your checkout page in a browser to see what WebMCP exposes. So it guesses, confidently.

MCP is the open standard that lets an assistant call a service's tools instead of guessing. With Orbilo connected, "is shop.example.com ready for agents?" becomes a real check: the assistant calls check_site, gets the score and grade per spec, every check that passed, warned or failed, and a fix for each. It answers from that, in the conversation you are already having, and hands you the report link.

The checker part works without signing in. That was deliberate. The public checker at /check has never needed an account and the MCP version should not either. If you want an assistant to be able to answer the question for any site, it can.

What the server exposes

Five tools and two resources.

• check_site checks a URL against UCP, WebMCP and Access, or one of them, and returns a score and grade per spec, the individual checks with fixes, and a shareable report link. No account needed. Runs live, so a full three-spec check can take up to a minute.
• get_report reads a report back by its link or token for 30 days. No account needed.
• list_monitored_sites returns your team's monitored sites with the latest grade per spec, how many issues are open on each, and how many site slots the plan has left. Needs an account.
• get_site_status is one site in detail: the latest stored check per spec, WebMCP per monitored page, and every open issue with its fix. Needs an account.
• add_monitored_site starts monitoring a site, runs the first checks straight away and uses one of the plan's site slots. Needs an account.
• orbilo://changelog/ucp and orbilo://changelog/webmcp carry the plain-English changelog so the assistant can explain a release rather than invent one.

Every tool is annotated so the assistant knows which are read-only and which change state, and the two public tools declare that they need no authentication, so a client does not send you to a consent screen just to run a check.

Use cases

These are the workflows we designed the tool set around.

Answer "are we ready?" from wherever the question comes up.

"Is shop.example.com ready for AI agents?"

check_site with the URL. The assistant reports the grade per spec and reads the failed checks to you. Because it has the fix for each, it can tell you what to do, not only what is wrong. That is the whole Learn section applied to your site in one reply.

Check one spec, fast.

"Check example.com for WebMCP only and tell me what is failing."

check_site with spec: webmcp. A single spec comes back far quicker than all three, and for WebMCP the assistant can see exactly which tools the page declares and which of them are broken.

Fix it with the codebase open.

In Cursor, VS Code or Claude Code the assistant already has your storefront's code in front of it. Ask it to "check staging.example.com for UCP and fix what fails." It runs the check, reads the failing checks and their fixes, edits the manifest or the JSON-LD, and when you have deployed, runs check_site again with fresh: true to skip the cache and confirm. The finding and the fix live in the same conversation.

Read a report someone sent you.

"Here is an Orbilo report link, what does it say?"

get_report with the link. Reports are shareable for 30 days, so a colleague, a client or an agency can hand one over and the assistant reads it back without a check being rerun.

Morning status on everything you monitor.

"Which of my monitored sites has open failures?"

list_monitored_sites, then get_site_status for the ones with open issues. You get the grades, the issues and the fixes in one message instead of a tab per site.

Start monitoring by asking.

"Start monitoring example.com."

add_monitored_site. The first checks run immediately and the site is on the daily schedule from then on. If the plan is out of site slots, the assistant is told so and can point you at billing rather than failing quietly.

Explain a spec release without a guess.

"What changed in the last UCP release and does it affect checkout?"

The assistant reads orbilo://changelog/ucp, which is the same plain-English changelog we publish, and answers from it. When a release lands, a connected assistant can say what it means for your site the same day.

What it will not do

Some boundaries were deliberate.

• No remove tool. An assistant can add a monitored site but cannot remove one. Removal stays in the dashboard.
• No account changes. The server cannot change your plan, touch billing or create, read or delete API tokens.
• OAuth only for monitoring. The endpoint refuses Orbilo API tokens. You approve an assistant once on a consent screen and it receives a token limited to the team you were working in.
• Daily limits apply to checks. Without an account, the limit is the public checker's, counted per network address. Signed in, it is higher. The message tells you when it resets.
• Rate limited. The server has its own limiter so a runaway agent loop cannot hammer the checker.

Tokens expire after an hour and refresh on their own for 90 days. Every connected assistant is listed under Connected assistants on the API tokens page, with a Disconnect button that revokes it immediately.

Getting connected

Every client is covered on the MCP page and in the docs: ChatGPT, Claude, Claude Code, Codex, Cursor, VS Code, Gemini CLI and Windsurf. For checks, the URL is all you need. For monitoring, sign in when the assistant asks.

If you have never used Orbilo, run the free checker first to see what a report looks like. Monitoring one site is free permanently, not as a trial, and that includes the MCP server.