AgentAvow Documentation
Guides for checking, verifying, gating, and badging AI-agent tools.
- How scoring worksAgentAvow answers one question: is this tool safe for your agent to connect to? We point a scanner at a tool — a GitHub repo, a package (npm, PyPI, crates), a Hugging Face model, a container image, a live MCP server, or an Agent Skill — and answer with one of three phrases and the one reason behind it. Under the answer sit the evidence: a signed, offline-verifiable 0–100 trust score, the adoption score, and the findings.
- Gate on the answerA score you don't act on is trivia. AgentAvow is built so a machine can read the answer and decide — block a risky tool, throttle an unproven one, or wave a Certified one through — in CI and at your agent's runtime. Every path below reads the same signed result you can recompute offline.
- Reading your scan scoreAgentAvow scans the tools, MCP servers, packages, and skills your AI agents connect to, and answers with one of three phrases — Safe to connect, Review before you connect, or Do not connect — plus the reason, backed by a signed 0–100 trust score you can verify yourself. This guide explains how to read a result.
- Behavioral sandboxThe static scan reads a tool. The behavioral sandbox runs it and reports what it actually did. The run is signed as its own dated observation, and that signed observation is also an input to the trust score, by fixed rules anyone can recompute (see How the sandbox moves the score).
- Run locally & in CIScan your own code without sending it anywhere — same engine, same score as a hosted AgentAvow scan, but offline. Built for the inner loop and for private repos.
- Add a trust badgeShow that your tool is safe, with proof. An AgentAvow badge shows your tool's current answer (Safe to connect, Review before you connect or Do not connect), its signed trust score (0–100) on a 10-segment bar, and its adoption on a small dial. Clicking it opens the full, verifiable report. It's free, needs no account, and refreshes on its own, so it never goes stale.
- Verify an attestationEvery AgentAvow scan result ships with a signed attestation — a JWS (JSON Web Signature, EdDSA/Ed25519, RFC 7515) over the verdict. Anyone can verify it offline, against our published keys, without trusting our servers. That's the whole point: *not a score you take on faith — a signature you can check.*
- MCP connectorAgentAvow's remote MCP server lets Claude, ChatGPT (and any MCP client) check whether a tool is safe before an agent connects to it. This page is for users and IT admins evaluating the connector.
- Auto-scan in Claude CodeMake "check it before you connect it" automatic. Three opt-in ways, all your own configuration. AgentAvow's MCP server never forces a scan, so nothing here can flag your setup in a directory review. You opt in; you stay in control.