polymarket-toolkit-mcp

MCP server exposing polymarket-toolkit's read-only Polymarket data CLI as Model Context Protocol tools for AI agents. No keys, no orders — data only.

Quick start

claude mcp add polymarket-toolkit -- npx -y polymarket-toolkit-mcp

Generic MCP client config:

{
  "mcpServers": {
    "polymarket-toolkit": {
      "command": "npx",
      "args": ["-y", "polymarket-toolkit-mcp"]
    }
  }
}

Tools (10, all read-only)

| Tool | What it answers | |---|---| | pm_profile | PnL + open positions snapshot for an address/username | | pm_activity | Recent trades for an address/username | | pm_brier | Prediction-quality (Brier) score from settled positions | | pm_pnl_check | Fee-inclusive PnL cross-check vs leaderboard | | pm_scan | Active markets ranked by 24h volume + spread | | pm_updown | Crypto up/down market fields for an event slug | | pm_leaderboard | Profit leaderboard snapshot | | pm_redeem_watchdog | Redeemable positions for an address (no signing) | | pm_v2_check | V2 CTF readiness diagnostics | | pm_rate_limits | Known Polymarket API rate limits (offline) |

Security model

• Every tool shells out to the toolkit's pm CLI (argv array, no shell) — the server can do exactly what the read-only CLI can do, nothing more.
• Inputs are allowlist-validated twice (MCP SDK layer + independent handler re-validation).
• Subprocess guards: 60s timeout, 2MB output cap, concurrency limit.
• No trading over MCP — order placement lives in the toolkit's separate opt-in executor/ package, human-wired only.

Full docs: docs/mcp.md