I built a remote MCP canary for opaque hosts — Claude listed 258 tools; ChatGPT called the sentinel

[sjh9714]

I built MCP Host Canary to pinpoint the last MCP boundary a managed host actually reached.

It creates a disposable 30-minute no-auth MCP server, records only server-observed protocol stages, and exports a privacy-minimized receipt. A tools/call request and the actual tool callback are separate observations; the receipt never infers an event the server did not receive.

Initial controlled receipts differed:

• Claude Web / limits: fetched the full 258-tool manifest twice; two explicit requests produced no sentinel_257 callback.
• ChatGPT Developer mode / limits: fetched all 258 tools and executed sentinel_257 three times.
• Claude Web / baseline: listed three tools and executed one direct canary_ping; the indirect request produced no extra callback.
• ChatGPT Developer mode / baseline: direct and indirect repetitions reached six actual callbacks in total.

These are observations from this server, not conformance grades or claims about either host's internal cause.

This complements a local inspector: the disposable URL is connected to the managed host itself, so the receipt shows the last boundary that reached the remote server. The pilot stores no tokens, prompts, tool arguments, request bodies, IP addresses, or raw User-Agent strings.

• Try the live beta
• Source, privacy boundaries, and known limitations
• Share a safe receipt from your own host

The beta is free, experimental, in memory on one instance, and a restart can remove active runs. Treat each temporary capability URL as a secret.

Which managed host stopped at which server-observed stage for you?