agentseal-mcp-intel

Security scanner for MCP servers. Check any MCP server for supply-chain threats before you install it -- directly from Claude, Cursor, or Windsurf.

The problem

MCP servers can read your files, run commands, and call APIs on your behalf. Before you install one, you should know:

• Does it exfiltrate data to external servers?

• Does it contain prompt injection or jailbreak payloads?

• Does it request excessive permissions?

• Has it been independently tested?

Most people install MCP servers blindly. This tool changes that.

Related MCP server: tooltrust-mcp

What it does

agentseal-mcp-intel connects your AI assistant to the AgentSeal security registry -- 3,400+ MCP servers each tested with 265+ adversarial security probes. Your assistant can look up any server, scan your environment, and flag risks before you install anything.

Quick start

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "agentseal": {
      "command": "npx",
      "args": ["-y", "agentseal-mcp-intel"]
    }
  }
}

Cursor

Add to .cursor/mcp.json (project) or ~/.cursor/mcp.json (global):

{
  "mcpServers": {
    "agentseal": {
      "command": "npx",
      "args": ["-y", "agentseal-mcp-intel"]
    }
  }
}

Windsurf

Add to ~/.codeium/windsurf/mcp_config.json:

{
  "mcpServers": {
    "agentseal": {
      "command": "npx",
      "args": ["-y", "agentseal-mcp-intel"]
    }
  }
}

Restart your AI tool. That's it.

Tools

search_registry

Find MCP servers by what you need. Results include trust scores so you can pick a safe option.

"Find me a PostgreSQL MCP server"
"Search for web scraping MCP servers with a score above 70"

check_server

Check the security of a specific MCP server before installing it. Accepts package names, slugs, or GitHub URLs.

"Is @modelcontextprotocol/server-github safe?"
"Check the security of https://github.com/crystaldba/postgres-mcp"
"Check mcp-server-postgres before I install it"

Returns:

• Trust score (0-100) with risk level

• Score breakdown across 5 dimensions (description safety, schema safety, capability risk, auth & permissions, stability)

• Security findings with severity ratings

• Install links and metadata

check_environment

Scan all MCP servers currently installed in Claude Desktop, Cursor, and Windsurf. Checks each one against the registry and groups results by risk level.

"Scan my installed MCP servers for security issues"
"Check my Cursor MCP setup"

Privacy: Only server names are sent to the API. Config file paths, environment variables, API keys, and credentials never leave your machine.

check_file

Analyze an AI config file for hidden security threats. Runs entirely locally -- zero data is sent anywhere.

"Check my .cursorrules file for threats"
"Scan CLAUDE.md for prompt injection"

Detects:

• Prompt injection and jailbreak patterns

• Invisible Unicode characters (zero-width joiners, Unicode tags, BiDi overrides)

• Data exfiltration instructions (URL variable interpolation, fetch + secrets)

• Base64-encoded payloads

• Unsafe auto-approve permissions

check_file only accepts known AI config file patterns. Requests for arbitrary files (e.g. /etc/passwd, ~/.ssh/id_rsa) are rejected.

submit_server

Submit an MCP server that's not in the registry for security scanning. AgentSeal fetches it, runs 265+ probes, and publishes results.

"Submit mcp-server-postgres for scanning"
"Scan https://github.com/org/repo and add it to the registry"

Requires explicit confirmation. Your AI assistant cannot submit servers without you approving it first.

Example session

You:  Check if the GitHub MCP server is safe to install.

Claude: [calls check_server("@modelcontextprotocol/server-github")]

      [AgentSeal] server-github  ~ HIGH -- 82/100
      ████████████████░░░░  82/100
      TypeScript . MIT . 18 tools . by modelcontextprotocol

      SCORE BREAKDOWN
        Desc Safety      ▪▪▪▪▪▪▪▪▪▪  95
        Schema Safety    ▪▪▪▪▪▪▪▪··  80
        Capability       ▪▪▪▪▪▪▪···  72
        Auth & Perms     ▪▪▪▪▪▪▪▪··  78
        Stability        ▪▪▪▪▪▪▪▪▪▪  100

      FINDINGS
        0 critical . 1 high . 1 medium

You:  Scan all my installed MCP servers.

Claude: [calls check_environment()]

      ENVIRONMENT SCAN -- 4 servers across 2 config files

      SAFE
        server-github          82/100  HIGH
        server-filesystem      78/100  HIGH

      REVIEW
        mcp-server-browserbase 54/100  MEDIUM  (1 finding)

      UNKNOWN
        some-custom-mcp        NOT IN REGISTRY

      ACTION REQUIRED
        Run check_server("some-custom-mcp") or submit it for scanning.

Privacy and data handling

• No telemetry. No usage tracking, no analytics, no background network calls.

• Credentials never leave your machine. Config parsing extracts only server names.

• check_file is 100% local. File contents are never transmitted.

• External content is isolated. API responses are wrapped in <agentseal:external> delimiters so your AI assistant treats them as data, not instructions.

• Source code is fully auditable. You're looking at it.

Configuration

Free vs Pro

Get a Pro key at agentseal.org/pricing.

Requirements

• Node.js 18 or later

• An MCP-compatible host: Claude Desktop, Cursor, Windsurf, or any MCP client

Development

git clone https://github.com/agentseal/agentseal-mcp-intel.git
cd agentseal-mcp-intel
npm install
npm run build
npm test          # 133 tests
npm run typecheck # strict TypeScript

How it works

Your AI assistant                agentseal-mcp-intel               AgentSeal API
     |                                  |                               |
     |-- "check postgres-mcp" --------->|                               |
     |                                  |-- GET /api/v1/mcp/intel/... ->|
     |                                  |<-- trust score + findings ----|
     |<-- formatted security report ----|                               |
     |                                  |                               |
     |-- "scan my .cursorrules" ------->|                               |
     |                                  |-- [local analysis only]       |
     |<-- findings (no network call) ---|                               |

The MCP server is a thin, auditable client. All scoring intelligence lives in the AgentSeal API. Local analysis (check_file) runs entirely on your machine with zero network calls.

Security

If you find a security vulnerability, please email hello@agentseal.org instead of opening a public issue.

License

FSL-1.1-Apache-2.0 -- Functional Source License 1.1.

You may use and modify this software, but you cannot offer it as a competing commercial service. After 4 years, the license converts to Apache 2.0. The AgentSeal registry, probe suite, and scoring engine are proprietary. See fsl.software for details.