MCP Warden

Audit-first MCP proxy for AI agents. MCP Warden scans client configs, wraps stdio MCP servers, logs tool calls, flags risky tools, and can append Agentbox-compatible events to an active run.

30-second demo

npx @swarmclawai/mcp-warden@latest policy init --write
npx @swarmclawai/mcp-warden@latest scan --config ./mcp.json
npx @swarmclawai/mcp-warden@latest install --client cursor --config ./mcp.json

Dry-runs are the default. Add --write only after reviewing the plan.

Commands

| Command | Purpose | |---|---| | mcp-warden scan | Find MCP servers in known client configs or --config | | mcp-warden wrap --name <server> -- <command...> | Proxy a stdio MCP server and log risk events | | mcp-warden install --client <client> [--write] | Produce or apply a wrapper rewrite plan | | mcp-warden policy init | Print or write mcp-warden.config.json | | mcp-warden report | Summarize Warden audit events | | mcp-warden help-agents | Print the machine-readable command catalog |

Every data-returning command supports --json and emits one JSON line on stdout.

Supported Clients

V1 discovery recognizes claude-code, cursor, cline, windsurf, codex, and explicit --config <path>.

Agentbox Integration

When AGENTBOX_RUN_DIR is set, wrap appends compatible mcp and risk JSONL events into the active Agentbox run.

Policy

MCP Warden warns by default. Blocking is opt-in through mcp-warden.config.json:

{
  "schemaVersion": 1,
  "mode": "audit",
  "blockRiskCodes": []
}