About this project
Evolver provides a self‑evolving engine for AI agents built on the GEP (Genome Evolution Protocol). It runs as a Node.js CLI tool that must be executed inside a Git repository. On each run it scans the local `memory/` directory for logs and signals, selects matching Genes or Capsules from a local GEP asset store, and emits a strict GEP prompt to stdout. The prompt can be consumed by host runtimes (e.g., OpenClaw, Cursor, Claude Code) or used manually. Evolution events are recorded in `events.jsonl` for auditability.
The tool offers three primary modes: a one‑off evolution run, a `--review` mode that pauses before applying changes for human confirmation, and a `--loop` daemon that repeats the process with adaptive sleep. It supports configurable strategy presets (`balanced`, `innovate`, `harden`, `repair‑only`) via the `EVOLVE_STRATEGY` environment variable.
Evolver integrates with major agent platforms through `setup‑hooks`. Commands such as `evolver setup‑hooks --platform=cursor` write hook files to the respective IDE or CLI, enabling automatic evolution during session start, file edits, or stop events. It also includes optional network features: a heartbeat to the EvoMap Hub, a skill store for downloading reusable skills, a worker pool for accepting network tasks, and asset publishing for sharing Genes and Capsules.
The engine is environment‑agnostic, using environment variables for configuration (e.g., `MEMORY_DIR`, `A2A_HUB_URL`, `EVOLVE_REPORT_TOOL`). It includes an operations module (`src/ops/`) for lifecycle management (start, stop, status, check) and protects core source files from being overwritten by autonomous agents. All core functionality works offline; hub connectivity only unlocks collaborative and sharing capabilities.
Evolver is designed for teams that need auditable, protocol‑constrained evolution of agent prompts and logs, and for environments that require deterministic changes backed by a full evolution trail. It is not intended for one‑off scripts, free‑form creative rewrites, or systems that cannot tolerate the protocol overhead.
Comments
0 Rating appears after 10 ratings
Sign in to join the discussion.