AgentAvow/AgentAvow

> Formerly AgentGraph. The signed attestation format, JWKS, and existing badges are unchanged.
AgentAvow answers "is this tool safe for my agent to connect?" for any tool, MCP server, package, or skill: **Safe to connect**, **Review before you connect**, or **Do not connect**, with the reason, over a **signed 0–100 trust score** you can recompute and verify offline and a separate adoption score.
MCP Server — Trust & Security for AI Agents
Check the security posture of any agent or tool directly from Claude Code:
See sdk/mcp-server/ for setup and full tool list.
Connect from Claude
**Claude Code plugin (recommended).** The **AgentAvow Trust** plugin is listed in the Anthropic plugin directory. It bundles the MCP connector, a /scan command, a skill that scans a server or package before Claude adds or installs it, a SessionStart hook that checks each MCP server you have configured the first time it sees it (warn-only, fail-open), and a PreToolUse gate that checks the result on file before each MCP tool call: it denies a call to a server that reads Do not connect and asks before a tool whose definition changed since it was checked. You can also install it straight from this repo:
Then try /scan npm chalk. The hook skips servers on localhost or a private network, withholds any URL whose path looks like it carries a secret, and strips credentials and query strings from a URL before it leaves your machine. What it reads and sends. Source: plugins/agentavow-trust/.
**Connector only.** In Claude (web or desktop), add AgentAvow from the connector directory. In Claude Code without the plugin:
Connect from your editor
The remote MCP server is live at **https://agentavow.com/mcp** (Streamable HTTP, no auth, read-only) — nothing to install.
**One-click:**
- **Cursor** — cursor://anysphere.cursor-deeplink/mcp/install?name=agentavow&config=eyJ1cmwiOiAiaHR0cHM6Ly9hZ2VudGF2b3cuY29tL21jcCJ9
- **VS Code** — vscode:mcp/install?%7B%22name%22%3A%20%22agentavow%22%2C%20%22type%22%3A%20%22http%22%2C%20%22url%22%3A%20%22https%3A//agentavow.com/mcp%22%7D
**Or add it manually.** Cursor (~/.cursor/mcp.json, or per-project .cursor/mcp.json):
VS Code (.vscode/mcp.json, or **MCP: Open User Configuration**):
For the local stdio server instead: pip install agentavow-trust.
Key Features
- **Free, anonymous scanning** — Point AgentAvow at any GitHub repo, MCP server, npm or PyPI package, or OpenClaw skill (or a wallet address that resolves to one) and get one of three answers back with its reason. No account, no install. Results cache for 1 hour; ?force=true re-scans.
- **Two scores: trust + adoption** — Every scan leads with one of three answers and its reason (Safe to connect / Review before you connect / Do not connect), then a **0–100 trust score** (with a trust tier as detail) and an **adoption score** built from real usage (downloads, stars, installs). The trust score is computed from the findings; five per-category subscores (secret hygiene, code safety, data handling, filesystem access, dependency health) are reported beside it as independent axes. Each finding carries a severity and points at the exact line or manifest entry.
- **Signed, verifiable attestation** — Each result ships with a **JWS attestation** (EdDSA / Ed25519, RFC 7515) over a canonical verdict (RFC 8785 JCS). Anyone can **recompute and verify it offline** against the public JWKS at agentgraph.co/.well-known/jwks.json — the score is a product, the signature is the proof under it.
- **Trust tiers → recommended limits** — Each trust score maps to a trust tier (verified → blocked) with a recommended execution posture (req/min, token budget, confirmation prompts) so a gateway or agent framework can act on it automatically.
- **Trust badge** — A one-line, shields.io-compatible **SVG badge** for your README that renders the repo's current answer and signed trust score and links to the full verifiable report. Served with open CORS and refreshed from the hourly scan cache, so it never goes stale.