Keel

Overview

A platform where AI agents write trading strategies in a typed language, with validation and backtesting before live execution on Hyperliquid.

Q&A

How does Keel prevent backtesting errors?

Keel uses a typed strategy language where common mistakes cannot be expressed. The validator checks every draft before anything runs and produces errors specific enough for the agent to fix. Mistakes around timestamp handling, data alignment, and time-series calculations are prevented at the language design level, rather than caught after the fact.

What ensures backtest and live trading produce the same results?

When a strategy is saved, it is compiled into an artifact and pinned to specific component versions, similar to how a lockfile pins project packages. This ensures the same artifact and data produce the same results every time it runs, whether in backtest or live trading.

Who controls the trading account when running strategies on Keel?

The user controls the trading account. Strategies run on the user's own Hyperliquid account, and the user's keys never leave their wallet. Keel never holds funds or has custodial control.

Why is there no language model in the execution loop?

No language model runs in the execution loop to ensure trading is not slow, inconsistent, or hard to audit. Language models are used only during strategy development—to search, compose, validate and backtest strategies. What actually trades is a compiled artifact that runs deterministically without any model involvement.

How many strategy components are available to build with?

The agent picks from 218 typed components and wires them together. These fall into 13 categories, with the largest being signal transforms (61) and indicators (56). Each component declares its input and output types, parameters with allowed ranges, a version and a changelog.