Open Source Radar — September 10: agents get their hands on physical space
Today's board finally leaves the terminal. Three of the five hottest repos are about agents producing geometry you can build, print, or move a wall through, and the other two are about giving a whole team — or a whole desktop — a shared agent setup.
Pascal Editor (TypeScript, 23,200 stars, MIT) — An open-source 3D building editor, built on React Three Fiber and WebGPU, that runs in the browser or from a local CLI and exposes the scene to agents over MCP. It's the mid-rise version of the "agent with a real machine" idea: walls, slabs, zones, and furniture are structured nodes an agent can read and edit rather than pixels to guess at, and projects live in a local database instead of someone's hosted design tool. Two published skills do the work — one for setting up and driving the editor safely, one that checks whether a piece of furniture actually fits a room and admits what it can't verify. Use it when you want an agent that can rearrange a floorplan and show you the export, not one that writes a paragraph about rearranging it.
text-to-cad (Python, 15,266 stars, MIT) — A library of eleven agent skills covering the whole path from a plain-language request to a manufactured object: model it, export STEP, STL, 3MF or GLB, preview it locally, check the print, slice it to G-code, and hand the job to a Bambu printer. It also covers the robotics side, with skills for URDF and SRDF robot descriptions and SDF simulation worlds, plus a step.parts lookup for off-the-shelf hardware. The reason it's hot is the shape of the thing — CAD files are text an agent can write and diff, and the skills are deliberately bounded, with a DfAM check that measures wall thickness and overhangs before you waste a spool. If your agent's output so far has been code and slides, this is the on-ramp to parts.
Tencent teamai-cli (TypeScript, 3,423 stars) — A command-line tool that keeps a team's skills, rules, MCP servers, and docs in one shared git repo and pushes them out to every member's local agent — Claude Code, Codex, Cursor, CodeBuddy, WorkBuddy, OpenCode, and more. The flow is a familiar one: admins publish, changes go through a merge request, and a session hook pulls the latest version into each developer's tools automatically. Roles and tags decide who syncs what. It's early and some context features are still beta, but it addresses the problem nobody has solved cleanly: agent configuration right now is per-person folklore, and this makes it reviewable code.
PI-Desktop (TypeScript, 1,998 stars, LGPL-3.0) — A local-first desktop workspace for coding agents, built on Electron with a Rust host core, that is currently in early preview and featured on Product Hunt. Bring your own model, point it at a local repo, and pick one of three gates: Agent just works, Plan waits for you to approve a frozen implementation plan, Goal waits for you to approve the outcome and lets the agent choose the route. Subagents run independent work in their own context windows, and skills, MCP servers, and plugins are user-installable. No account and no mandatory relay — the pitch is the review layer, with privileged actions passing through a permission gate and diffs and command output sitting in a panel you actually look at.
awesome-gpt-image-2 (JavaScript, 30,539 stars, MIT) — A prompt library for image generation, now past 530 documented cases with full prompts and generation records, 20-plus reusable industrial templates, and installable skills. The current work is a GPT Image 2.5 section that puts the same prompt through the new Sunburst and Flare variants side by side, with a draggable comparison and four real recreations that keep the original gallery prompt. The authors flag that the original generation conditions and exact model IDs are unverified and that the site adds no live generation — honest caveats, rare in this genre. Read it as a reference for how to write image prompts that hold up across model versions, not as a gallery to copy.
Worth watching this week: the interesting frontier is no longer what an agent can say, it's which file formats it can write — STEP, G-code, URDF are the new API surface.
Would you trust an agent to send a part to a printer without a human checking the file first? Tell us in the comments.
Sources: Pascal Editor (GitHub) · Pascal Editor docs · text-to-cad (GitHub) · text-to-cad docs · teamai-cli (GitHub) · teamai-cli usage guide · PI-Desktop (GitHub) · PI-Desktop on Product Hunt · awesome-gpt-image-2 (GitHub) · GPT Image 2.5 spotlight
