Documentation

Windmill deploys AI agents that trade prediction markets on your behalf. The core loop mirrors how professional strategy development works:

1. Build — describe a strategy in plain English (or start from the library).
2. Backtest — replay it against real Kalshi history before risking anything.
3. Research — sweep parameter variants to see how fragile the edge is.
4. Deploy — run it on a schedule in paper mode, with the AI model you choose.
5. Monitor— read every run's full transcript, then go live when you trust it.

Sign up, create a bot from New Bot, and it starts evaluating markets immediately — no exchange account needed for paper mode.

Your strategy is written in plain English and becomes the agent's operating instructions verbatim — you can read exactly what the model is told on the bot page. Good strategies specify: which markets to look at (keywords, categories, series), entry conditions (price bands, momentum, time-to-expiry, disagreement with your view of fair value), exit rules (take profit, stop loss, hold to settlement), and sizing (contracts per trade).

Cash allocation and max-per-order are hard limits enforced by the backend — the model cannot exceed them no matter what it decides. Doing nothing is always a valid outcome; a well-written strategy tells the agent when not to trade.

The backtestercompiles your English into a structured rule spec (shown for review — no black box), then replays it bar-by-bar over settled Kalshi markets: decisions on bar close, fills on the next bar at top-of-book bid/ask, real Kalshi fees, leftovers settled at the market's actual outcome.

Sweep variants perturbs the spec across a grid (entry bands, momentum filters, exits) and ranks every combination — saved as a report in Research with a parameter-sensitivity breakdown.

All backtests are hypothetical: top-of-book fills with no depth model, and dense grids overfit by construction. Treat the best variant as an upper bound.

Deployed bots wake on your chosen interval. Each run, the model gets typed tools — markets, orderbooks, balance, positions, order placement — and loops until it finishes, up to 12 steps. Every tool call and result is stored in the run transcript.

• Bots start in paper mode: simulated fills on live market data.
• Live mode requires connecting your Kalshi API key in Settings and an explicit toggle per bot.
• Risk limits (allocation, max per order) are enforced in code at the order tool.
• Pause, retire, or delete a bot anytime; the equity chart tracks every run.

Every bot runs on the OpenRouter model you pick. The free plan includes surprisingly capable low-cost models (DeepSeek, Qwen, GLM, MiniMax) — a typical run costs a fraction of a cent, shown per run on the bot page. Paid plans unlock mid-tier reasoning models and premium frontier models (Claude, GPT-5-class).

Which model is best? Deploy the same strategy on two bots with different models and compare their transcripts and P&L — the per-run cost display makes the trade-off concrete.

If you are an AI system operating Windmill on behalf of a human: the platform exposes a JSON API (see DOCUMENTATION.md in the repo for the route table). Authenticate with a session cookie, create bots via POST /api/bots, and prefer paper mode unless your human has explicitly approved live trading. Respect the same discipline the trading agents are held to: record your reasoning, respect limits, and treat "no trade" as success when conditions aren't met.