polymarket-paper-trader

polymarket-paper-trader

[PyPI] [PyPI downloads] [Tests] [ClawHub] [License: MIT]

A zero-risk gym for AI agents on real prediction markets — real order books, official fees, verified fill fidelity. Practice, evaluate, and benchmark decision intelligence.

Agents make probability judgments all day. Polymarket is the world's largest prediction market, and its order books are the honest scoreboard: real money, real prices, real outcomes. But you cannot hand an agent a wallet to learn with. So this project gives every agent what SWE-bench gave coders — a faithful environment where judgment has consequences and gets scored:

• Practice — your agent trades $10k of paper money against live Polymarket order books, with the same fee model and fill mechanics as the real exchange
• Evaluate — the polymarket-benchmark harness in this repository (installed separately) scores any model on prediction-market decision sets (Brier score, calibration, alpha)
• Compare — multi-account battles and leaderboards rank agents against each other
• Compete — Forecast Arena: the live public leaderboard where AI models and simple baselines forecast the same Polymarket questions daily and are scored by real outcomes (Brier score). Runs from this repo (enter your model · workflow · raw data).

Part of agent-next — building an agentic world.

60-second demo

pip install polymarket-paper-trader
pm-trader init --balance 10000                 # $10k paper money
pm-trader markets search "bitcoin"             # find markets
pm-trader buy will-bitcoin-hit-100k yes 500    # buy $500 of YES
pm-trader stats --card                         # shareable stats card

Prefer not to install? uvx --from polymarket-paper-trader pm-trader-mcp runs the MCP server your agent connects to (see Works with below).

That's it. Your AI agent is now trading Polymarket with zero risk.

Install

# via pip (CLI + MCP server)
pip install polymarket-paper-trader

# or the MCP server alone, with no install
uvx --from polymarket-paper-trader pm-trader-mcp

# from source (development)
uv pip install -e ".[dev]"

Requires Python 3.10+.

Works with

Full copy-paste config for every runtime above (plus Grok/xAI remote MCP): docs/integrations.md. Hosts that implement MCP Apps render portfolio and stats as an inline dashboard; every other host gets the same text result.

Use it from ChatGPT / Claude.ai / personal agents

Remote mode exists for hosts that only speak HTTP: run pm-trader-mcp --transport streamable-http with the opt-in --auth-issuer, --auth-audience and --resource-url flags, then add your server's /mcp URL in the host's remote-MCP settings. It is off by default, and it needs an OAuth/OIDC issuer you operate — pm-trader verifies bearer tokens against it and gives every verified subject its own paper accounts. There is no hosted endpoint to add: you run the server yourself. Setup, threat model and the environment equivalents: docs/remote-mcp.md.

Quick start

# Initialize with $10k paper balance
pm-trader init --balance 10000

# Browse markets
pm-trader markets list --sort liquidity
pm-trader markets search "bitcoin"

# Trade
pm-trader buy will-bitcoin-hit-100k yes 100      # buy $100 of YES
pm-trader sell will-bitcoin-hit-100k yes 50       # sell 50 shares

# Check portfolio and P&L
pm-trader portfolio
pm-trader stats

How it works — and why to trust it

• Your order walks the real book. A buy consumes live ask levels from the lowest price upward, exactly like a real taker order — slippage is real and reported in basis points
• Fees follow the official per-match curve — fee = C × rate × p × (1-p), rounded to 5 decimals, makers exempt (exact spec in the CHANGELOG) — charged per filled level from each market's published fee schedule, not an approximation
• Paper cash, real discipline. Resting buys reserve their cash, partial fills keep their remainder open, closed or paused markets reject trades, and limit prices are validated against tick size
• Resolution pays $1/share. Call resolve (or resolve --all) when a market closes and winners pay out like the real thing
• Verified fidelity. The live test suite asserts that simulated fills land inside the band of prices the market actually quoted (Data API v2 price history) and that fees match the official curve exactly — run against real APIs on CI
• Upstream-aligned. The client tracks the current Gamma / CLOB / Data API surface, contract-verified by live probes
• Multi-outcome markets — any number of outcomes, not just YES/NO
• Kalshi too. Paper trade Kalshi markets with the same commands and tools by passing a kalshi:<ticker> slug (pm-trader buy kalshi:KXNBAGAME-26OCT16DENLAL-LAL yes 25); find them with kalshi markets. Fees follow each series' published curve (docs/kalshi.md)
• 100% coverage gate on the core package, plus end-to-end tests against the live API

CLI commands

Global flags: --data-dir PATH, --account NAME (or env vars PM_TRADER_DATA_DIR, PM_TRADER_ACCOUNT).

MCP server — what your agent can do

Your agent gets the following tools via the Model Context Protocol. The server also carries the full trading playbook with it — as MCP server instructions, as a trading_playbook prompt, and as a skill://trading-playbook resource — so MCP-only clients (Cursor, Claude.ai connectors, Grok API remote MCP, ChatGPT apps) get the same guidance skill-aware agents get from skill/polymarket-paper-trader/SKILL.md.

pm-trader-mcp  # starts on stdio

# or, with no local install:
uvx --from polymarket-paper-trader pm-trader-mcp

Add to your Claude Code config:

{
  "mcpServers": {
    "polymarket-paper-trader": {
      "command": "pm-trader-mcp"
    }
  }
}

Remote transport (streamable-http)

For MCP clients that only speak HTTP, run the server with --transport streamable-http:

pm-trader-mcp --transport streamable-http --host 127.0.0.1 --port 8000
# or: pm-trader mcp --transport streamable-http --host 127.0.0.1 --port 8000

The MCP endpoint is then http://<host>:<port>/mcp. There is no isolation between callers — everyone who reaches the server shares all of its paper accounts; it is self-host-only, not a public multi-user service. There is no authentication on this transport; anyone who can reach the port can call every exposed tool. backtest and pk_battle (local file reads + local strategy-module execution) are stdio-only and are not registered when serving over streamable-http.

Docker

docker build -t pm-trader-mcp .
docker run -p 127.0.0.1:8000:8000 -v pm-trader-data:/root/.pm-trader pm-trader-mcp

Publish the port to 127.0.0.1 only (as above) unless you put a real authenticating proxy in front of it — the container has no auth of its own.

MCP tools

Strategy examples

Three ready-to-use strategies in examples/:

Momentum (examples/momentum.py)

Buys when YES price crosses above 0.55, takes profit at 0.70, stops loss at 0.35.

pm-trader strategy run examples.momentum.run

Mean reversion (examples/mean_reversion.py)

Buys when YES price drops 12+ cents below 0.50 fair value, sells when it reverts.

pm-trader strategy run examples.mean_reversion.run

Limit grid (examples/limit_grid.py)

Places a grid of limit buy orders below current price with take-profit sells above.

pm-trader strategy run examples.limit_grid.run

Jev edge (examples/jev_edge.py)

"Jev vs the market": asks Jev for a YES probability per binary market and buys the side it favors, skipping markets priced outside [0.05, 0.95] where fees dominate — pm-trader strategy run examples.jev_edge.run (needs pip install -e "benchmark").

Writing your own strategy

Strategies are imported from the examples. package (the allowlist lives in pm_trader/benchmark.py), so drop your file there:

# examples/my_strategy.py
from pm_trader.engine import Engine

def run(engine: Engine) -> None:
    """Your strategy receives a fully initialized Engine."""
    markets = engine.api.search_markets("crypto")
    for market in markets:
        if market.closed or market.yes_price < 0.3:
            continue
        engine.buy(market.slug, "yes", 100.0)

pm-trader strategy run examples.my_strategy.run

For backtesting with historical data:

def backtest_strategy(engine, snapshot, prices):
    """Called once per historical price snapshot."""
    if snapshot.midpoint > 0.6:
        engine.buy(snapshot.market_slug, snapshot.outcome, 50.0)

Evaluate your agent: polymarket-benchmark

The paper trader is the gym; the polymarket-benchmark package in this repository is the scoreboard. It is a separate install (not part of the pm-trader CLI — pm-trader strategy replays trading strategies (pm-trader benchmark remains as an alias), see the CLI table above):

pip install -e "benchmark[dev]"
cd benchmark && polymarket-benchmark run --model opencode/jev-1.13-free --market-set mini

Run two models head-to-head to see whose judgment is actually better. Market sets, scoring (Brier, calibration, alpha) and model setup: benchmark/README.md.

Multi-account support

Run parallel strategies with isolated accounts:

pm-trader --account aggressive init --balance 5000
pm-trader --account conservative init --balance 5000

pm-trader --account aggressive buy some-market yes 500
pm-trader --account conservative buy some-market yes 100

pm-trader strategy compare aggressive conservative

Share your results

Generate a shareable stats card and post to X/Twitter:

pm-trader stats --tweet    # X/Twitter optimized
pm-trader stats --card     # markdown for Telegram/Discord
pm-trader stats --plain    # plain text

AI agents can use the stats_card MCP tool to generate and share cards automatically.

Honest limits

• Paper only. No wallet, no keys, no real trades, no real money — ever. Resolution payouts are simulated $1/share
• Simulation quality is verified against live order books and price history, but real execution adds queue position, latency, and counterparty behavior no simulator can promise
• Live market data needs network access to Polymarket's public APIs (no key required)

OpenClaw / ClawHub

Available on ClawHub as polymarket-paper-trader:

npx clawhub install polymarket-paper-trader

GitHub bot

Comment /oc or /opencode on an issue or PR. New issues get a triage reply; non-draft PRs get a shallow review. Implementation starts only if a human adds bot:implement or comments /oc implement (FreeInference deepseek-v4-flash; the implement push uses GITHUB_TOKEN scoped to contents/issues/PRs only — no Actions access — so a separate dispatch-tests job, the only job holding actions: write, re-validates the pushed branch ref, verifies .github/workflows/test.yml on that ref is identical to the default branch, then runs gh workflow run Tests --ref — token pushes do not trigger CI). Merging is always manual: the bot never merges, and required checks are enforced by branch protection. Q&A / triage / review stay on qwen3.6-35b. No wallet, no real trades. Sessions are not shared.

Also in this repository

The paper-trader is the product; three companion packages live alongside it.

See CONTRIBUTING.md for how to work on each package and CHANGELOG.md for release history.

Tests

pytest -m "not live"             # unit + integration, 100% coverage gate
pytest                           # full suite (requires network)
pytest tests/test_e2e_live.py    # live API integration tests only

License

MIT