Documentation
How to install chi, what its moving parts are, how to define a problem, and what the security model guarantees.
chi (χ) is an open-source autoresearch harness: it points a fleet of LLM coding agents at any problem that exposes a programmatic evaluator — a build, a correctness check, and a score — and iterates unattended.
These pages cover the practical surface. For the full design rationale, see the
specs in the repository’s docs/ directory.
Getting started
Install chi, open a session, configure providers, and run your first autoresearch loop.
Concepts
The director, the fleet, the blackboard store, the negative ledger, the watchdog, and steering — chi's moving parts and why each exists.
Problems & evaluators
Anatomy of a problem pack: problem.yaml, evaluator entrypoints, held-out seeds, two-tier evaluation, and fleet strategy.
Security
The threat model, sandbox tiers, and why submission gating holds even under full autonomy.
Changelog
What shipped in each chi release.