jac-workbench

JAC is a desktop AI workbench, an Electron app in TypeScript and Python, that started as a VS Code extension and grew into a standalone app. An agent can search, code, run tools, design 3D parts and circuit boards, and talk back, in one app. It works with any model on OpenRouter or a local model through llama.cpp or Ollama, gives it real tools through the Model Context Protocol, and adds purpose-built workspaces for CAD, PCB design and voice. The Windows installer auto-updates, and downloads live in the separate jac-desktop-releases repository.

The agent loop streams text, runs tools and feeds results back for up to 50 steps per turn, recovers when a reply is cut off or the model announces an action without taking it, and compacts long conversations into a Goal, Done, Learned, Failed, Open summary instead of failing on context overflow. Built-in tools include web search and fetch, an offline Wikipedia dump, full-text recall over the app's own saved conversations, read-only GitHub access through the user's own gh CLI, shell, and screenshots. MCP servers use a hand-written stdio JSON-RPC client with a multi-server manager, and every tool is permission-gated (allow, ask, deny). Voice is hands-free with a "hey Jac" wake word and two Whisper tiers.

CAD is a three.js viewer with move, rotate and scale gizmos backed by a build123d geometry server: the model writes the geometry code, JAC runs it and renders the part. PCB is a schematic editor with a component library, automatic footprint placement, A* auto-routing and KiCad export. The README's demo cases are a 40 by 20 by 3 mm bracket with two M3 holes and a 25 mm flange, and an ESP32-S3 LED driver board placed, wired into four nets, auto-arranged and auto-routed.