Waiting
Enter an MCP endpoint to see results.
MCP Server Check
Check a remote MCP server from outside before connecting it to an agent: authentication, tools that act, tool-poisoning text in descriptions, CORS and HTTPS. No tool is called.
Detected
Tools (name · kinds that act · annotations)
—
JSON
—
Tool Features
Related
Check an MCP server before you connect it
A remote MCP server gives an AI agent tools. Before you connect one, it helps to know whether it asks for credentials, which of its tools act on the world (run commands, write files, send messages, move money), and whether its descriptions try to steer the agent — the "tool poisoning" seen in published cases. This check reads what the server itself publishes and never calls a tool.
What we send
An initialize request and tools/list, one more initialize with a foreign Origin header (to see whether a web page on another site could call it), and a request for the OAuth protected-resource metadata of RFC 9728. The server instructions and every tool and parameter description are read with the same patterns we use for reported servers. Query strings are removed from the address before anything is sent or stored.
FAQ
Does a clean result mean the server is safe?
No. It means nothing was found in what the server shows from outside. What a tool does when called, the server code and its dependencies, and tools behind a login are not checked — the result lists them every time.
Will this run the server's tools?
No. Only the handshake and the tool list are requested. No tool is ever called.
Can I check a local MCP server?
No. Only public addresses can be checked; private, loopback and local addresses are refused.