AgentGuard — security checks for AI agents (MCP server)

A free, open-source, single-binary MCP server that gives any AI agent (Claude Desktop, Cursor, Claude Code, …) security checks in the loop:

No API key, no account, no telemetry. Everything runs locally except CVE lookups, which hit the free public OSV.dev API.

Install

Pick whichever fits your setup. All three give the same stdio MCP server.

Docker (no Go toolchain needed)

docker run -i --rm ghcr.io/shuaicongxiaomai/agentguard:latest demo

MCP client config (mcpServers block in Claude Desktop / Cursor / Claude Code):

{
  "mcpServers": {
    "agentguard": {
      "command": "docker",
      "args": ["run", "-i", "--rm", "ghcr.io/shuaicongxiaomai/agentguard:latest"]
    }
  }
}

Go install

go install github.com/shuaicongxiaomai/agentguard@latest

{ "mcpServers": { "agentguard": { "command": "agentguard" } } }

Prebuilt binary

Download the archive for your OS/arch from the Releases page, extract it, and point the config command at the absolute path to the binary.

Related MCP server: opzyai

Build from source

go build -o agentguard .          # append .exe on Windows

Three ways to run it

1. demo — see it work, zero config:

agentguard demo

Runs all three tools on sample inputs and prints the results.

2. stdio (default) — local use; the MCP client launches the binary. No port; the client manages its lifecycle. Use any of the configs above, then ask your agent: "scan this for secrets", "is lodash 4.17.20 vulnerable?", "check my package.json for vulnerable deps."

3. serve — Streamable HTTP service on a port (for a hosted/remote server):

agentguard serve :8080          # or set PORT=8080
# health:   GET  http://host:8080/healthz
# MCP URL:  POST http://host:8080/mcp   (clients connect to this URL)

Use stdio for local use; use serve when you want a hosted server users connect to by URL without installing anything.

Notes

• scan_secrets runs locally and never returns raw secret values — only masked previews.

• scan_dependencies caps at 200 deps per call and queries OSV concurrently.

• check_cve / scan_dependencies need network access to OSV.dev; scan_secrets is fully offline.

License

MIT.