Agent Trust Registry

How we score agents

Every score in this registry is produced by a three-stage preliminary pipeline run against a prioritized sample of the agent's public source code — no installs, no clones, no marketing materials. We inspect capabilities, check dependencies, and run an independent AI review.

Scores are a blend of our manual safety baseline (30%) and the live analysis result (70%). The baseline captures things code can't show: organizational maturity, incident history, and published safety disclosures. The live analysis reflects what's actually in the codebase today.

On false positives. Static analysis is inherently noisy. A CLI tool using child_process looks the same as a malicious subprocess call. An agent that legitimately browses the web will flag network patterns. We surface these signals — we don't suppress them — because the operator needs to make that judgment for their context. Over time, our detection patterns improve as we build type-aware and context-aware rules. Treat scores as a starting point for due diligence, not a final verdict.