Your AI agent's prompts, but with version control and a PhD
A Node.js CLI that turns ad-hoc prompt tweaks into auditable, reusable "genes" instead of forgotten markdown files.

What it does
Evolver is a prompt generator, not a code patcher. It scans a ./memory/ directory for logs and signals, selects a compact “Gene” or “Capsule” from its local asset store, and emits a strict GEP (Genome Evolution Protocol) prompt to stdout. You copy that prompt into your agent, or let a host runtime like OpenClaw consume it. Every cycle writes an EvolutionEvent for audit. It runs fully offline; git is mandatory for rollback and blast-radius tracking.
The interesting bit
The project ships with an arXiv paper (2604.15097) backing its core bet: across 4,590 trials, compact “Gene” representations outperformed longer documentation-style “Skill” packages and proved more robust under structural perturbation. The README is unusually frank about what the tool does not do — no automatic code edits, no arbitrary shell execution, and --loop is explicitly “not a real-time agent assistant.” That restraint is rarer than it should be in this space.
Key highlights
npm install -g @evomap/evolver, thenevolverinside any git repo — 30-second onboarding- Hooks for Cursor, Claude Code, Codex, Kiro, opencode, and native OpenClaw integration
- Three modes: single run,
--review(human confirmation), or--loop(background daemon) - Optional EvoMap network connection for skill sharing and leaderboards; core engine works offline
- Research-backed: cites controlled trials showing gene-evolved systems lifting base-model performance on CritPt physics tasks
Caveats
- Codex integration has a documented gap: no session transcript file, so
--reviewfalls back toMEMORY.md,USER.md, or a self-generatedmemory_graph.jsonl - License shifting: future releases will move from GPL-3.0 to source-available after a perceived uncredited clone appeared in March 2026; existing MIT/GPL versions remain freely usable
Verdict
Worth a look if you’re already managing agent prompts across multiple projects and want auditability without building your own governance layer. Skip it if you need an agent that actually edits code autonomously — this is strictly a prompt-evolution engine.