New open-source coding agent 'Forge' beats Devin on SWE-bench
A new open-source AI coding agent achieves 54% on SWE-bench — surpassing Devin's published scores — and runs locally on a MacBook Pro.
What happened
A team of researchers released Forge, an open-source coding agent that achieves 54% on SWE-bench Verified (Devin's last published score was 45.8%). Forge runs locally via a CLI, uses any OpenAI-compatible API or local Ollama models, and is available on GitHub under MIT license. It uses a tree-search approach for exploring solution paths and has built-in GitHub PR creation. Works with Claude, GPT-4, and Llama 3 backends.
Why it matters to you
personalizedWhy it matters to you
An open-source coding agent that runs locally, beats Devin, and works with any OpenAI-compatible API is significant for three reasons: (1) You can use it today, free. (2) You can study and modify the architecture — tree-search for code editing is novel and worth understanding. (3) Local execution means no code leaves your machine, which matters for proprietary codebases. The MIT license means you can build products on top of it.
What to do about it
Install Forge today: `pip install forge-agent`. Point it at a medium-complexity bug in your codebase and watch how it explores solutions. Even if it doesn't fully solve the task, studying its approach will improve how you prompt other coding tools. If it works well, integrate it into your CI pipeline for automated issue triage.
Tags
Sources