Five Open Source Projects That Supercharge Claude Code

Author: helloahad_

A roundup of five recently released open source projects designed to extend Claude Code's capabilities, from automated ML experimentation to self-improving agent skills to full Google Workspace access.

AutoResearch by Karpathy

The headliner of the list. AutoResearch brings ML-style experimentation loops into Claude Code. You point it at something you want to improve, say a Python app's runtime performance, and it runs iterative experiments to find optimizations. In the example shown, AutoResearch ran 83 experiments and kept 15 improvements. The catch: the target needs a measurable, binary outcome. Speed benchmarks work well. Subjective quality does not.

OpenSpace: Self-Improving Skills

OpenSpace watches how you use Claude Code skills and sorts them into three improvement buckets. Auto-fix kicks in when a skill breaks. Auto-improve activates when the skill works but could follow better patterns. Auto-learn captures winning workflows from real usage. The project's authors report a 45% reduction in token consumption and a 4.2x increase on the GDP economic benchmark.

CLI Anything

From the same team behind OpenSpace. CLI Anything lets you point Claude Code at any open source project and generate a CLI tool for it. The pitch is that MCPs are losing ground to CLIs, and this tool bridges that gap. They have already used it to build CLIs for GIMP, Blender, and Inkscape.

Claude Peers

This project enables Claude Code sessions running in separate terminals to communicate with each other. One session generates code while another evaluates it, creating a tight feedback loop. Useful for harness-style workflows where you want generation and validation running in parallel.

GWS: Google Workspace CLI

An unofficial CLI built by Google developers that gives Claude Code full access to your Google Workspace suite, including Docs, Sheets, Drive, and the rest. Not a Google product, but built by people who know the APIs.

Key Takeaways

• AutoResearch works best when your optimization target has a clear, measurable outcome like execution speed or accuracy
• OpenSpace's three-bucket approach (fix, improve, learn) offers a structured path to skill evolution without manual tuning
• CLI Anything reflects a broader trend: CLIs are replacing MCPs as the preferred integration layer for AI coding tools
• Claude Peers opens up multi-agent patterns where one instance writes and another reviews
• GWS fills a major gap for anyone whose workflow depends on Google Workspace