The state of agent tool safety.
An independent index of signed answers for the tools, MCP servers, and packages AI agents connect to — checked from the outside, no repo access, verifiable offline. Not the agent you build; the things it trusts.
Where the risk is
Share of checked tools with a critical finding, by kind
Critical findings over tools checked, per kind. Counted in catalog rows, not unique repos.
From the sandbox
last 30 daysBeyond reading the code, we run packages and MCP servers in an isolated sandbox, call every tool, and plant canary credentials. In the last 30 days: 11,650 runs, 1,117 servers exercised, 203 behavioral findings.
Aggregate counts only. Each run is signed as a separate observation shown beside the score. How the sandbox works →
The leaderboards
Ranked from the live catalog — click any tool for its full signed report.
Safest MCP servers
by trustMost relied upon
by adoptionSafest npm packages
by trustSafest PyPI packages
by trustCheck the tool you're about to connect.
Paste any repo, package, or MCP server — get an answer and a signed, offline-verifiable trust score in seconds. Free, no account.
