# AgentAvow/AgentAvow [Health: Active]

**Category:** 🔒 Security  
**Repository:** https://github.com/AgentAvow/AgentAvow  
**GitHub Stars:** 5  
**Views:** 0  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/agentavow-agentavow

## Description
Signed, recomputable safety scores for MCP servers, packages and tools, with offline-verifiable attestations and a GitHub Action to gate CI merges.

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `uvx` (confidence: high):

```json
"mcpServers": {
  "agentavow": {
    "command": "uvx",
    "args": ["agentavow-trust"]
  }
}
```

## Documentation

## What AgentAvow/AgentAvow does

AgentAvow evaluates software and agent-facing resources before an AI agent connects to them. It accepts GitHub repositories, MCP servers, npm packages, PyPI packages, OpenClaw skills, and wallet addresses that resolve to one of those resources. Each scan produces one of three decisions: Safe to connect, Review before you connect, or Do not connect.

The result includes a trust score from 0 to 100, a separate adoption score, and the reason for the verdict. Trust is broken down across secret hygiene, code safety, data handling, filesystem access, and dependency health. Individual findings include severity and a reference to the relevant source line or manifest entry.

## How it works

The hosted MCP endpoint is available at `https://agentavow.com/mcp` over Streamable HTTP. It is read-only and does not require authentication. Results are cached for one hour; a forced scan can bypass that cache with `force=true`.

Each result includes a JWS attestation using Ed25519 signatures and RFC 7515 EdDSA. The verdict is canonicalized with RFC 8785 JCS, allowing consumers to recompute the score and verify the signature offline against the published JWKS. Trust tiers map scores to recommended execution limits, including request rates, token budgets, and confirmation prompts.

## Setup and configuration

For a hosted connection, configure an MCP client with the remote URL. Cursor and VS Code provide one-click installation links, and both also support manual MCP configuration. Claude web or desktop can add AgentAvow through its connector directory; Claude Code can connect with `claude mcp add --transport http agentavow https://agentavow.com/mcp`.

The local connector is installed with:

```bash
pip install agentavow-trust
```

The repository also supports Docker Compose and local development. Self-hosting uses Python 3.11 or newer, Node.js 20 or newer, PostgreSQL 16, and Redis 7. Docker Compose starts the backend, frontend, PostgreSQL, and Redis services, with database migrations applied during startup.

## Tools and capabilities

- Scan repositories, MCP servers, packages, and skills before connection.
- Return trust verdicts, scores, adoption data, tiers, and detailed findings.
- Provide signed attestations for offline verification.
- Watch resources for score changes and signed tool-definition changes.
- Expose a trust badge for repository documentation.
- Gate GitHub, GitLab, Bitbucket, Azure Pipelines, or local builds when a scan returns Do not connect.

The Claude Code plugin adds a `/scan` command, a skill for checking servers or packages before installation, a SessionStart check for configured MCP servers, and a PreToolUse gate based on the stored result. Its checks are warn-only and fail-open at session start, while tool calls can be denied when the result is Do not connect or confirmation is needed after a tool-definition change.

## Limitations and notes

The hosted service scans anonymously and exposes read-only operations. Its plugin skips localhost and private-network servers during its automatic checks, avoids sending URLs whose paths appear to contain secrets, and removes credentials and query strings before sending a URL for scanning. Private repository scans can use a GitHub token supplied transiently; the token is not persisted or added to the public catalog.

The public catalog contains launch data and community-triggered scans. Watch alerts depend on rescanning and compare signed tool-definition digests to detect changes that may otherwise go unnoticed.

_Full upstream README: https://allmcps.com/mcp/agentavow-agentavow/readme_