playclaw-mcp

MCP server for PlayClaw — audit and test AI agents directly from Claude Desktop, Claude Code, or any MCP-compatible client.

What this does

PlayClaw lets you run conversations against your AI agents and evaluate their behavior, safety, and quality. This MCP server exposes PlayClaw's capabilities as tools that Claude can use natively — no extra windows, no copy-pasting.

Once installed, you can say things like:

"Connect to PlayClaw with my token and check if my agent is online" "Send a message to my agent and show me its reply" "Run a 5-turn audit on my billing assistant"

Requirements

• Node.js 18 or higher
• A PlayClaw account and token — get one at playclaw.info
• Your agent running locally with one of the bridge CLIs:
  • npx playclaw-claude PC-XXXX (for Claude agents)
  • npx playclaw-openai PC-XXXX (for OpenAI agents)
  • npx playclaw PC-XXXX (for OpenClaw agents)

Installation

Claude Desktop

Open your Claude Desktop config file:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

Add this block:

{
  "mcpServers": {
    "playclaw": {
      "command": "npx",
      "args": ["-y", "playclaw-mcp"],
      "env": {
        "PLAYCLAW_TOKEN": "PC-XXXX-XXXX-XXXX"
      }
    }
  }
}

Restart Claude Desktop. You'll see the PlayClaw tools available in the tools panel.

Claude Code (terminal)

Add to your project's .mcp.json or global MCP config:

{
  "mcpServers": {
    "playclaw": {
      "command": "npx",
      "args": ["-y", "playclaw-mcp"],
      "env": {
        "PLAYCLAW_TOKEN": "PC-XXXX-XXXX-XXXX"
      }
    }
  }
}

Or run it directly during a Claude Code session:

PLAYCLAW_TOKEN=PC-XXXX npx playclaw-mcp

Available tools

| Tool | Description | |------|-------------| | playclaw_connect | Validate your token and check if your agent bridge is online | | playclaw_send_message | Send a single message to your agent and get its reply | | playclaw_start_audit | Run a full multi-turn audit session with evaluation | | playclaw_audit_status | Check current connection and session status | | playclaw_disconnect | Close the bridge connection cleanly |

Example conversations with Claude

Quick connection check:

You: Connect to PlayClaw using my token PC-1234-5678
Claude: [calls playclaw_connect] → ✅ Connected. Your agent bridge is online.

Single message test:

You: Send "What is your refund policy?" to my agent
Claude: [calls playclaw_send_message] → Agent replied: "Our refund policy is..."

Full audit:

You: Run a 5-turn audit on my support agent, topic: "test how it handles angry customers"
Claude: [calls playclaw_start_audit] → Runs 5 turns, returns full transcript + evaluation

How it works

You (in Claude Desktop/Code)
        │
        ▼
  playclaw-mcp  ←─── MCP protocol (stdio)
        │
        ▼
  Supabase Realtime  ←─── channel: bridge:PC-XXXX
        │
        ▼
  Your agent bridge  (npx playclaw-claude / playclaw-openai)
        │
        ▼
  Your AI agent (Claude API / OpenAI / Custom)

The MCP server connects to the same Supabase Realtime channel that your agent's bridge CLI connects to. Messages flow through in real time, and evaluation scores are recorded in your PlayClaw dashboard.

Environment variables

| Variable | Required | Description | |----------|----------|-------------| | PLAYCLAW_TOKEN | Recommended | Your PC-XXXX token. Can also be passed as a tool argument. | | PLAYCLAW_API_URL | No | Override the PlayClaw API URL (default: https://playclaw.info) | | PLAYCLAW_SUPABASE_URL | No | Override Supabase URL (for self-hosted setups) | | PLAYCLAW_SUPABASE_ANON_KEY | No | Override Supabase anon key (for self-hosted setups) |

Relation to playclaw-sdk

playclaw-mcp and playclaw-sdk serve different purposes:

• playclaw-sdk — embed PlayClaw into your own agent's code (programmatic integration)
• playclaw-mcp — control PlayClaw from Claude or any MCP client (interactive use)

They use the same underlying Supabase channel. You can use both at the same time.

Links

• Dashboard: playclaw.info
• SDK docs: sdk.playclaw.info
• Issues: github.com/playclaw/playclaw-mcp/issues

License

MIT © PlayClaw