How the index checks servers
If you saw mcpwatch-index in your logs, this page explains what it did and how to opt out.
What we send
Once a day, for each remote endpoint a publisher listed in the official MCP registry, the indexer sends the same requests any MCP client sends when it connects:
initialize(the standard MCP handshake)notifications/initializedtools/list(reads the tool names, descriptions and input schemas)- an HTTP
DELETEwith the session id, to close the session
That's all. We never call a tool, never send credentials, never guess passwords and never fuzz. If your endpoint asks for authentication, we stop and record "requires auth".
Limits we keep
- At most 3 endpoints per server and at most 2 concurrent connections per hostname.
- A 20-second budget per endpoint and a 1 MB response cap.
- Only public internet addresses: private, loopback and cloud-metadata addresses are refused, and redirects are not followed.
- User-Agent:
mcpwatch-index/0.2 (+https://www.mcpwatch.io/probe).
Opt out
Open an issue titled "Opt out: <your registry server name or hostname>" at github.com/anardimontlv/mcpwatch. Opted-out servers are no longer contacted. Their page keeps only the checks based on public registry data and says "not probed (owner opt-out)"; opting out never lowers a grade.
Why we do this
Developers connect AI agents to MCP servers, and an agent reads tool descriptions as instructions. Publishing which tools a public endpoint exposes, and noticing when they change, helps people decide what to trust. The method is documented in how it works, and the indexer's source code is public.