Start

Run a check

Everything you can do on the check form, how to read the result and what to do with it.

#The connector URL

Paste the full address of your MCP endpoint, the same address you would give Claude under Add custom connector. It must start with https://. The endpoint is usually a path such as /mcp; the address of your marketing site or docs will not work, because it does not answer MCP requests.

The check sends real requests to that address from our server on the public internet. A localhost address, a private network address or a name that only resolves inside your office or VPN fails at the DNS step, which is exactly what Claude would do.

#Servers that use a fixed token

If your server takes a static API key instead of OAuth, tick My server uses a fixed token or API key. Two fields appear:

FieldWhat to enterWhat is sent
HeaderThe header your server reads. Leave empty for Authorization. Letters, numbers and dashes only.The header name you typed.
TokenYour key or token.With Authorization, Bearer <token> unless you already typed Bearer, Basic or Token in front. With any other header, the value exactly as typed.

#Reading the verdict

The big headline summarises the whole run. There are three verdicts:

VerdictMeaning
failAt least one step failed. The headline names the first failing step: Claude would stop at: Public DNS. Fix that one first; steps after it were skipped because they depend on it.
warnNothing fails, but one or more steps found something that can break connections now and then, such as a certificate close to expiring or a slow first response.
passEvery step we can test from the public internet passes.

Step statuses

StatusMeaning
passThe step did what Claude needs.
warnIt works, but with a risk. The summary says what, and the fix says how to remove it.
failClaude would stop here. This is the step to fix.
skipNot run, with the reason. Either an earlier step failed, or the step does not apply (a server without sign-in has no sign-in steps).

Each step also lists its evidence: the raw facts MCPatch saw, such as the DNS records, the certificate issuer and expiry, the response status and headers, and the first part of a response body. That is what you would otherwise have to collect with curl and dig. The report also shows how long each step took and, when it can tell, which host runs the server (Vercel, Cloudflare, Render, Railway, Fly.io, Netlify, AWS, Azure, Google Cloud, nginx or an ngrok tunnel).

#The fix prompt

Under the steps, Copy fix prompt gives you text written for an AI coding assistant. Choose the tab that matches your tool: Claude Code and Cursor get instructions to find and change the code in the repository, ChatGPT and Other get a version that asks which files it needs to see first. Every prompt includes:

  • your server URL and the host, if detected;
  • each failing or warning step, in the order Claude hits them, with what happened, the evidence and the required fix;
  • the connector rules that apply (Claude's egress range, IPv4 only, HTTPS only, S256 PKCE, the redirect address, the 10 second limits);
  • a request for curl commands to verify each fix from outside your network.

When everything passes but Claude still fails

Sometimes every step passes from outside and Claude still says Couldn't reach. Then the problem is likely on Anthropic's side. The fix-prompt box becomes a support report: a formatted summary of what passed, ready to paste into a new issue on the connector repository, together with the ofid_ reference ID that appears in the URL of Claude's error.

#Sharing a report

Every check gets a report at /r/<id>. Anyone with the link sees the same result, which makes it easy to send to a teammate, a vendor or a support thread. Tokens are never part of a report. Reports are kept so the link keeps working, until you ask us to delete them: email the link to the address in the privacy policy.

#What else runs with a check

Three extra results appear below the twelve steps. They are explained on Extra checks:

  • Other MCP clients: which protocol versions your server accepts, because other clients ask for different versions than Claude.
  • Tool changes: what changed in your tool list since your last check of the same connector.
  • Hidden instructions: a scan of tool descriptions for text aimed at the AI.

#Limits

LimitValueWhat you see
Free checks per visitor3 per 24 hoursThe free-limit message, with the plans.
Free checks per connector3 per 24 hours across all visitorsThe same message. A new network address does not reset it for the same server.
Request rate12 per 10 minutes per addressHTTP 429, Too many checks. Try again in a few minutes.
Install checksNot counted against the free connector checksSame request rate limit as above.

Paid plans re-check your connectors every day without using these limits; see Dashboard and plans.

#What the check cannot see

  • It does not run from Anthropic's network. MCPatch checks from its own server on the public internet. If your firewall only admits Anthropic's egress range (160.79.104.0/21), Claude gets through but MCPatch will be blocked, and the report will say so at the reachability step.
  • It cannot sign in for you. For OAuth servers it tests every step up to and including the token endpoint with a deliberately invalid code, but it never completes a real sign-in.
  • It tests what Claude does at connection time. It does not run your tools.
A passing result in one line
https://mcp.example.com/mcp  [Vercel]  1340 ms
Everything we can test from the public internet passes