Geneclaw Documentation

Everything you need to install, configure, and safely evolve your agents. Start with Getting Started or jump to any topic.

Core Concepts

What is Geneclaw?

Geneclaw is an open-source framework for safe, controlled agent self-evolution. It provides a structured workflow — the Geneclaw Evolution Protocol (GEP) — that allows agents to observe their own behavior, diagnose failure modes, propose improvements as structured diffs, and apply those improvements only after passing a multi-layer safety gate and receiving explicit human approval.

Geneclaw is built on top of nanobot (HKUDS/nanobot), which provides the underlying agent execution and tool-use runtime. Geneclaw adds the evolution, safety, auditing, and reporting layer on top.

Key Design Principles

  • Dry-run by default: No change is ever applied without an explicit --apply flag. Every command is safe to run in read-only mode.
  • Human approval required: The Gatekeeper can auto-approve proposals in autopilot mode only within explicitly configured risk thresholds. Outside those thresholds, human review is mandatory.
  • Everything is audited: All events, proposals, gate decisions, and apply results are written to an append-only JSONL event store with secret redaction.
  • Rollback is always possible: Changes are applied on git branches with automatic rollback if tests fail. Manual rollback is one command away.
  • Containment by default: The Gatekeeper's allowlist limits what paths the agent can touch. Start narrow; expand deliberately.

Data Layout

Geneclaw uses a minimal, transparent file structure:

  • data/events.jsonl — append-only event store (all agent events)
  • proposals/ — generated GEP JSON proposals
  • benchmarks/ — benchmark results by cycle
  • geneclaw.toml — main configuration file