what is trustcard fr?

Trustcard is a security tool for a growing situation where AI agents connect to and run tools built by other people, such as MCP servers, plugins, and remote services, without ever having inspected them directly. It answers a basic question: before an agent calls one of these tools, how does it know what the tool actually does, who approved it, whether it has changed since last time, and whether this specific call should be allowed. The tool works by turning each capability, such as an MCP server's set of tools, into a unique cryptographic fingerprint based on what it actually exposes. Publishers can then digitally sign a manifest describing that exact set of tools, so a client can later confirm that what it received matches what was signed. Trustcard also tracks trust over time, moving a connection through states like unknown, observed, pinned, and, if something changes unexpectedly, mismatch or revoked, so a later connection tells you not just that the server responds, but that it is still the same capability you trusted before. When something does change, trustcard classifies how serious that change is, distinguishing something harmless like a wording tweak from something risky like a new destructive action appearing or a safety restriction being removed. At the moment a tool is actually called, trustcard applies two checks: first, is this the capability that was approved, and second, is this particular call, with these particular arguments, allowed under policy. It can block calls if a server has been revoked, no longer matches its trusted identity, includes an unrecognized tool, or attempts a destructive action against policy. The project also includes a separate scanner that inspects a live MCP server across eight checks, such as whether it can be installed, whether it correctly speaks the MCP protocol, and whether its tool schemas are valid, producing a practical health score. Trustcard can be installed globally or run without installation via npx, and can also be used as middleware placed between an MCP client and server in JavaScript code. The full README is longer than what was shown.