OpenEvolve 🤖 Agent Open source

What it does

OpenEvolve is an open-source, evolutionary coding framework inspired by DeepMind's AlphaEvolve. You give it starting code and an evaluation function, and it uses an ensemble of LLMs to iteratively mutate and improve that code across generations, aiming to autonomously discover better algorithms rather than just answer one-off coding prompts.

What is inside

• An island-based architecture running several evolutionary populations in parallel
• Support for an LLM ensemble spanning OpenAI, Google Gemini, local models, and the Claude Code CLI as a generation backend
• An "artifact side-channel" that feeds execution errors and outputs back into the next generation as feedback
• MAP-Elites quality-diversity optimization to keep a diverse set of candidate solutions instead of collapsing to one
• Multi-objective Pareto optimization for balancing competing goals
• Full-run reproducibility via comprehensive seeding (default seed=42) and real-time evolution-tree visualization

Works with

Designed to call out to multiple LLM providers as its mutation engine, explicitly including the Claude Code CLI alongside OpenAI, Gemini and local models.

How to install or connect

pip install openevolve

A Docker image is also published at ghcr.io/algorithmicsuperintelligence/openevolve:latest, and a development install from source is documented.

Maintenance and safety

Apache-2.0 licensed, with roughly 7,400 stars, 825 commits and 73 open issues at review time — an active research-adjacent project rather than a dormant fork. Running it costs real money per iteration (the project estimates roughly $0.01–$0.60 depending on provider and code complexity), since each generation makes LLM calls, so budget and rate-limit accordingly.

Who should use it

Engineers and researchers who want to automate discovery of better algorithms or code (e.g. numerical methods, heuristics, scheduling logic) rather than just get one agent-assisted patch, and who are comfortable writing an evaluation/scoring function and monitoring iteration spend.