What it does
ToolTrust Scanner analyses MCP tool definitions before an agent uses them and assigns each one a trust grade from A to F. It runs 16 static tool-definition rules and 2 source-scan rules for embedded MCP implementations. The checks cover prompt injection hidden in descriptions, over-broad permissions, scope mismatches, known CVEs in bundled dependencies via OSV, privilege escalation, and arbitrary code execution patterns. It also flags confirmed compromised packages, such as specific LiteLLM, Trivy and Langflow versions. It is an MCP server and a CLI/CI check, not a host, gateway or runtime proxy.
It suits developers who want to audit the MCP servers in their setup before installing them, or who want a pass/fail check in a CI pipeline. The agent can run the config scan itself, which reads the MCP configuration, connects to each server in parallel and returns a risk report with grades and enforcement decisions. In a labelled research snapshot of 207 servers, 70% had at least one finding and 10% received a clean Grade A; the publisher's directory holds current, directory-wide results.
It runs over stdio, so there is no network listener: the host launches it as a subprocess. Dependency lookups are sent to the OSV service, and the scanner needs access to it to produce a grade. The publisher states that skills and other agent tool formats are planned but not yet covered.
Limitations and requirements
If an OSV lookup fails, including through rate limiting, service errors, malformed responses or timeouts, the scan reports 'scan incomplete' and issues no grade; the CLI exits nonzero and the gate does not install the server. An OSV outage can therefore prevent scanning of tools with declared dependencies. Declared dependency names and versions are sent to the OSV endpoint, and an optional TOOLTRUST_OSV_URL override must be trusted and use HTTPS. There is no fallback to the public API when an override fails. Skills and additional agent tool formats are not yet covered.
Safety
Our team has not reviewed this resource. Only automated rules checked it; resources with high-risk signals are not published. These are signals, not a guarantee that installing it is safe. Try it in a sandbox project first.
- Runs code or changes an assistant's behaviour
Sources
- Official MCP Registry first seen 13 Nov 2026 · last seen 13 Nov 2026 · its page on the source
ToolTrust Scanner
Scans MCP servers for prompt injection, data exfiltration, and privilege escalation.
v1.0.9As shown on registry.modelcontextprotocol.io when we last read it