Five Open Source Projects That Supercharge Claude Code (March 2026)
Watch on TikTok
Five open source projects dropped in March 2026 that significantly extend what Claude Code can do -- from autonomous ML experimentation to self-improving skills to controlling desktop applications from the terminal. Here is what each one does and why it matters.
1. AutoResearch by Andrej Karpathy
AutoResearch brings machine learning experimentation into Claude Code. Point it at something you want to improve -- say, a Python app you want to run faster -- and it runs experiments continuously, keeping only the changes that beat the current best result. In one example, it ran 83 experiments in succession and kept 15 improvements. The constraint: you need a measurable, binary metric (faster/slower, better/worse) for the system to evaluate against.
The project is 630 lines of Python, runs on a single GPU, and has accumulated over 21,000 GitHub stars since launch.
2. OpenSpace (Self-Improving Skills)
OpenSpace treats skills as living entities rather than static config files. Point it at a task, and it evaluates your skill usage across three buckets:
| Bucket | What It Does |
|---|---|
| AutoFix | Repairs skills when they break |
| AutoImprove | Optimizes working skills with better patterns |
| AutoLearn | Captures winning workflows from actual usage |
The results from real-world testing: 45% reduction in token consumption and a 4.2x increase in income generated using the GDPVAL economic benchmark.
3. CLI-Anything
CLI-Anything lets you point Claude Code at any open source project and generate a CLI tool for it. This is significant because as MCPs fall by the wayside, CLI tools are taking their place as the preferred integration pattern. The project ships with pre-built CLIs for over 30 applications including GIMP, Blender, Inkscape, Audacity, OBS Studio, and LibreOffice.
4. Claude Peers
Claude Peers enables your Claude Code sessions to communicate with each other. Run a terminal in one window and another terminal in a second window, and those sessions can discover each other and exchange messages in real time. The practical use case: one Claude Code session generates code while a second evaluates it, creating an instant feedback loop. This pattern -- planner, executor, evaluator -- maps directly to how engineering teams operate.
5. GWS (Google Workspace CLI)
GWS is a Google Workspace CLI that gives Claude Code complete access to your Google suite -- Drive, Gmail, Calendar, Sheets, Docs, Chat, Admin, and more. It is not an official Google product in the traditional sense, but it is created and maintained by Google developers under the googleworkspace GitHub organization. Google ships updates when their APIs change, which solves the maintenance problem that plagues third-party integrations.
Key Takeaways
- AutoResearch automates iterative ML experimentation with measurable improvement loops
- OpenSpace makes skills self-improving, reducing token costs by 45%
- CLI-Anything is replacing MCPs as the preferred way to give agents control over desktop software
- Claude Peers enables multi-session collaboration patterns (generator + evaluator)
- GWS gives Claude Code native access to the full Google Workspace, maintained by Google
Resources
- AutoResearch -- Autonomous ML experimentation by Andrej Karpathy
- OpenSpace -- Self-evolving skill engine for AI agents
- CLI-Anything -- Generate CLIs for any open source software
- Claude Peers -- Inter-session communication for Claude Code
- GWS CLI -- Google Workspace CLI with AI agent skills
Published April 18, 2026. Writeup generated from a favorited TikTok.