# evolver

> CLI that turns agent logs into auditable prompts built from Genes and Capsules

- Repository: https://github.com/EvoMap/evolver
- Category: Agent skills & plugins
- Stars: 8,697 · Forks: 842 · Language: JavaScript · License: GPL-3.0
- Experience: Some experience
- Page: https://ir.adyog.com/repos/evomap/evolver/

Teams that run AI agents often fix problems by editing prompts by hand, and nobody can later say what changed or why. Evolver replaces that with a fixed loop. It reads runtime logs and error patterns from a memory/ folder, then picks a reusable asset called a Gene or Capsule that fits the problem.

It then prints a prompt in a format called GEP, the project's own protocol, and writes an EvolutionEvent to disk as an audit record. Evolver is a prompt generator, not a code patcher. It does not edit your source files or run arbitrary shell commands. You copy the prompt into your agent, or a host such as OpenClaw, Cursor or Claude Code picks it up through hooks.

The README argues that compact Genes beat longer skill documents and cites the authors' own paper (arXiv 2604.15097). That is the authors' claim, so weigh it as such. The core runs offline, and an optional EvoMap network connection adds skill sharing and leaderboards. The README also says future releases will become source-available, while already published GPL-3.0 versions stay usable.

## What it does

- **Gene and Capsule selection**: Scans the memory/ directory for logs and signals, then chooses the best-matching asset from a local GEP asset store.
- **Audit trail**: Each run writes an EvolutionEvent into ./memory/, so you can trace what was suggested and when.
- **Review and loop modes**: `--review` pauses for human confirmation. `--loop` runs as a background daemon for self-maintenance tasks.
- **Agent runtime hooks**: `setup-hooks` wires Evolver into Cursor, Claude Code, Codex, Kiro and opencode. OpenClaw reads its stdout directives directly.
- **Works offline**: Core functions need no internet. It does require Node.js 18 or newer and a git repository, which it uses for rollback and blast radius calculation.

## What you could build

- Generate a structured improvement prompt from your agent's recent error logs.
- Keep an auditable history of how an agent's behavior was changed over time.
- Hook Evolver into Cursor or Claude Code so each session produces evolution events.
- Review each proposed evolution step by hand before it reaches your agent.
- Run a background loop that maintains agent memory and assets on a schedule.

## Who it's for

Suits teams or developers who already run AI agents with logs and want traceable prompt changes; you should be comfortable with Node.js, git and the command line.

## Get started

Install the package globally, then run `evolver` from inside a git-initialized project. It scans ./memory/, prints a GEP prompt to stdout and records an EvolutionEvent; copy the prompt into your agent or set up a hook for your runtime.

```
npm install -g @evomap/evolver
```
```
evolver --help
```
```
evolver
```

Tags: agent-evolution, gep, prompt-governance, cli, skills