How patchbay scans public MCP servers
Last updated 2026-11-19.
patchbay runs a passive security auditor against public Model Context Protocol (MCP) servers. This page explains what we scan, how we scan it, and what we never do without consent.
What we scan
- OAuth 2.1 Authorization Server Metadata (RFC 8414):
/.well-known/oauth-authorization-server - Protected Resource Metadata (RFC 9728):
/.well-known/oauth-protected-resource - MCP
tools/list,resources/list,prompts/listvia the MCP JSON-RPC endpoint, without authenticated calls - Transport-security headers on discovery endpoints (HSTS, X-Content-Type-Options, Referrer-Policy, Permissions-Policy, Content-Security-Policy)
- TLS certificate metadata read from the TLS handshake
We read only what a browser or a casual integrator would see. We follow the host's robots.txt if present. Scanner timeout per host is 30 seconds. Rate limit per host is 1 request per second.
What we never do
- We never call authenticated tools.
- We never attempt credential enumeration, brute-force, or fuzzing.
- We never upload or infer private data.
- We never store request payloads beyond what is needed to produce a finding record.
- We never retain scan traffic longer than 30 days.
How findings are labeled
- source-confirmed — we observed the issue directly in a response the server returned to us. Reproducible with curl.
- runtime-verified — we observed a behaviour across repeated requests that confirms the condition.
- potential-fp — the signal is present but the context is ambiguous. Marked as such.
No finding is ever presented as a confirmed exploit. A finding is a signal to check.
Public scope
Our scan set includes public MCP server hosts that are listed on the public MCP registry, linked from a published Protected Resource Metadata document, or added via public submission.
If your server is listed and you want it removed from the public scan set, email sarah@patchbay-mcp.dev with the host name and a brief statement that you are the operator. We will remove it within 48 hours. We publish a withdrawn set for retracted hosts so the public history stays honest.
Why this is lawful
We scan only public endpoints the host advertises for discovery or that are explicitly made available. We do not bypass authentication. We do not exceed what a casual integrator checking compatibility would do. Our legitimate interest basis under GDPR Article 6(1)(f) is documented at /legitimate-interest-assessment.