agent-audit

Security scanner for AI agents and MCP servers. Find vulnerabilities in 30 seconds.

Two modes: scan your local codebase for agent security issues, or audit a remote MCP server for tool-level risks.

Quick start

# Scan local codebase
npx @authora/agent-audit

# Scan a remote MCP server
npx @authora/agent-audit mcp https://mcp.example.com

# With authentication
npx @authora/agent-audit mcp https://mcp.example.com --api-key YOUR_KEY

MCP server scanning

Connects to any MCP server, discovers all tools, and classifies each one by risk level.

npx @authora/agent-audit mcp <url> [options]

Options:

| Flag | Description | |------|-------------| | --api-key <key> | Authenticate with an API key (sent as api-key header) | | --bearer <token> | Authenticate with a Bearer token | | --json | Output raw JSON (for CI pipelines) | | --fail-below <grade> | Exit with code 1 if grade is below threshold (A+, A, B+, B, C, D) |

Examples:

# Scan with API key authentication
npx @authora/agent-audit mcp https://mcp.authora.dev --api-key authora_live_xxx

# JSON output for CI
npx @authora/agent-audit mcp https://my-server.com --bearer sk-xxx --json

# Fail CI if grade drops below B
npx @authora/agent-audit mcp https://my-server.com --api-key xxx --fail-below B

Sample output:

  MCP Security Audit  via REST /tools
  --------------------------------------------------

    B   Grade  65/100  AUTH

  128 tools  63 safe  45 review  20 dangerous

  Dangerous tools:
  * authora_suspend_agent -- Suspend an active agent
  * authora_revoke_agent -- Permanently revoke an agent
  * authora_delete_role -- Delete a role by ID
  ...

  Needs review:
  * authora_create_agent -- Create a new Authora agent in a workspace
  * authora_update_agent -- Update an existing agent
  ...

  README badge:
  ![MCP Security: B](https://img.shields.io/badge/MCP_Security-B-yellow)

Tool classification:

| Level | Keywords | Meaning | |-------|----------|---------| | Safe | get, list, search, verify, audit, health, describe... | Read-only operations | | Needs Review | create, update, assign, configure, send, rotate... | State-changing mutations | | Dangerous | delete, revoke, suspend, destroy, kill, force, reset... | Destructive or privilege-escalating |

Scoring:

Score is based on the ratio of safe vs dangerous tools, not absolute counts. A server with many tools but few dangerous ones scores well. Authentication adds 10 points.

| Grade | Score | Typical server | |-------|-------|----------------| | A+ | 90-100 | Read-only analytics, monitoring | | A | 80-89 | Well-scoped API with auth | | B+/B | 60-79 | Full CRUD platform with auth | | C | 50-59 | Broad API, no auth | | D/F | 0-49 | Mostly destructive, undocumented |

Local codebase scanning

Scans your project files for agent security issues.

npx @authora/agent-audit [directory] [options]

Options:

| Flag | Description | |------|-------------| | --json | Output as JSON | | --badge | Generate README badge markdown |

What it checks:

| Category | What it finds | |----------|--------------| | Credentials | Shared API keys across agents, hardcoded secrets in code | | Identity | Missing agent identity layer, no cryptographic verification | | MCP | MCP servers without authentication, unprotected tool endpoints | | Permissions | Overly broad agent permissions, admin/root access | | Delegation | Missing delegation chains, agents inheriting full user permissions | | Audit | No audit logging for agent actions | | Approvals | No human-in-the-loop for sensitive operations | | Resilience | Missing timeouts, no error handling on tool calls |

Sample output:

  Agent Security Audit
  by Authora -- https://authora.dev

  Scanning current directory...

  Scanned 47 files
  Found 3 agent(s), 2 MCP server(s)

  CRITICAL  Shared API key may be used by 3 agent files (.env)
  CRITICAL  No agent identity layer detected
  WARNING   MCP server without visible auth configuration (mcp/server.ts)
  WARNING   No delegation chains

  Agent Security Score: 1.5/10  Grade: F

CI integration

# GitHub Actions
- name: MCP Security Gate
  run: npx @authora/agent-audit mcp ${{ secrets.MCP_URL }} --api-key ${{ secrets.MCP_KEY }} --fail-below B

Web inspector

Prefer a browser UI? Use the MCP Security Inspector -- same scanner, visual interface.

Learn more

• awesome-agent-security -- curated resources
• Authora -- identity, coordination, and security for AI agents
• Authora MCP docs -- securing MCP with Authora

License

MIT