Know which of your MCP servers and dependencies are safe before your AI uses them.
At your next session start, your AI opens with one line: how many of your servers AgentAvow checked, how many are Safe to connect, and the one to look at first, plus the same for your project's direct dependencies. Add a server with `claude mcp add`, or install a package with npm, pnpm, yarn, bun, pip, uv, pipx or poetry, and it is checked first: Do not connect asks you, Review or deprecated is a one-line note, Safe
Example path → ·
https://agentavow.com/mcpAgentAvow grades the safety of anything an AI agent connects to — a GitHub repo, an MCP server, an npm/PyPI/crates/Docker/Hugging Face package, or a wallet-linked identity — and returns one of three answers with its reason (Safe to connect, Review before you connect, or Do not connect), a signed 0-100 trust score that anyone can recompute offline, and an adoption score.
When to use which tool:
— a GitHub repo, passed as 'owner/name'.
— a published package (registry + name).
— a live MCP server, by its https URL (checks the tool definitions themselves for poisoning/injection).
/ / / — agent-to-agent identity and trust.
Reading a verdict: each result leads with one of three answers and its reason: 'Safe to connect', 'Review before you connect', or 'Do not connect', with ' · Certified' where the tool earned it. A 0-100 trust score and an adoption score sit underneath. Every result links to a full report and a signed (Ed25519/JWS) attestation. All tools are read-only and need no account.
A good habit is to scan a tool before connecting or installing it. To do that automatically for every new tool, a user can add a one-line CLAUDE.md rule or a SessionStart hook (setup: https://agentavow.com/docs/auto-scan-claude-code) — this is the user's own opt-in config.
1 in the Claude listing