MCP/Risk

Open security scoring

Should you connect
that MCP server?

try

Trust grade and a separate capability rating for any Model Context Protocol (MCP) server, against a published, open-source policy. It reports what an outside observer can check, and does not certify that a server is safe.

Beta. Policy v0.2, checks still being tuned. Read a grade as a reason to look closer.

What the checks look at

§AreaMethod
§1ProvenanceRegistry listing, repository health, package and repo consistency, npm provenance attestation.
§2CapabilitiesTool surface read via tools/list or static source analysis. Untrusted code is never executed. Shell, filesystem, egress, and credential access are flagged.
§3Transport & authHTTPS, authentication required, OAuth 2.1 protected-resource metadata (June 2025 spec).
§4VulnerabilitiesKnown advisories via OSV.dev, where a checkable package exists.
§5Instruction integrityHidden instructions, zero-width characters, cross-tool shadowing, and authority over-reach across tool descriptions, initialize instructions, prompts, and resources. Schema changes between scans are flagged as a rug-pull signal.

Browse the directory