What does the MCP security checker test?
It connects to your server the way an AI client does, with an MCP “initialize” request. It checks HTTPS and the certificate, whether a login is required, your OAuth discovery documents and PKCE support, CORS and HSTS headers, and version leaks. If the server is open, it lists the tools and flags risky ones without safety hints, prompt-injection phrases in tool descriptions, and whether requests from foreign websites are refused.
Is the check safe for my server?
Yes. It sends a handful of requests, never calls a tool, never sends passwords or tokens, and closes its session when done. Only check servers you own or are allowed to test.
My server needs a login. Can you still check it?
Yes, and that is the best case. We check that strangers get a 401 reply, that the reply tells AI clients where to log in, and that your login server supports PKCE with S256 over HTTPS, as the MCP specification asks. We don’t look at the tools behind the login.
Is an MCP server without login always unsafe?
No. A server that only shares public, read-only information can be open, like public documentation. It becomes risky when open tools can change data, send messages or reach private systems. The checker tells you which case you are in.
What is tool poisoning?
It is a prompt-injection attack hidden in a tool’s description, such as “ignore previous instructions” or a request to read secret files. AI clients read these descriptions, so hidden orders can change what the AI does. The checker looks for common phrases; a manual review finds more.
Does it replace a security review?
No. It checks the outside of your server in seconds. A full review also tests what each tool can reach, permission scopes, logging and real prompt-injection attempts.