Open Source Radar — August 17: agents go desktop-native
Today's board is light on new coding agents and heavy on the layers around them — making the desktop software agents can't yet reach runnable, making their diagrams look like design work instead of tool output, and making fine-tuning something you can do on a laptop. Three fresh projects worth your attention.
CLI-Anything (Python, ~47.7k stars) — A project out of HKUDS that wants to make every piece of software agent-native: it wraps desktop tools — CAD packages, Blender, Obsidian, LibreOffice, game engines — into clean, agent-callable command-line interfaces, then catalogs them in a browsable hub where you install community-built harnesses for whatever app you live in. The star count is the signal: roughly 47.7k in a few months, because it attacks the exact gap agents still choke on — code and APIs are easy for them, GUI software isn't. If your agent can't reach the desktop tools your actual work lives in, this is the bridge.
diagram-design (HTML, ~20k stars) — An agent skill for Claude Code, Codex, and Pi that produces editorial-grade diagrams: self-contained HTML and SVG, 27 visual types from architecture to flowcharts to pyramids, no Figma, no generic rounded boxes, no "Mermaid-slop." The pitch is that agent-generated diagrams have all blurred into a sameness nobody wants to publish, and this one matches your site's brand in about a minute by reading it. If your agents keep handing you diagrams your designer can't look at, this is the fix — use it whenever the output needs to read like design work, not tool output.
Soup (Python, ~2.1k stars) — A one-command fine-tuning tool: one YAML config sets up training, with auto-detection of GPU, quantization, and batch size, and no SSH into a rented box or hours of config hunting. Its standout trick is layer streaming, which keeps the frozen base model out of VRAM so you can LoRA-fine-tune an 8B model on a 4GB laptop GPU — measured around 119 tokens/sec peaking near 3.3GB with results bit-identical to a normal run. It's honest that the trick is beta, and it even ships a release gate that won't call a tune an improvement unless the gain clears the model's own noise floor. If fine-tuning has always meant renting cloud GPUs, this one removes the excuse.
Worth watching this week: all three point the same way — the next frontier isn't smarter agents, it's making the software, output, and training around them work for ordinary setups.
Would you hand your laptop's desktop apps to an agent, or keep them human-first? Tell us in the comments.
Sources: CLI-Anything (GitHub) · diagram-design (GitHub) · Soup (GitHub)
