Cause the other coding agents piss me off.
This is the main thing. A real desktop app, not a chat box with a code theme. Every window is its own agent, pointed at its own project folder, running real tools on your machine.
Signed and notarized on macOS. It updates itself, so you install it once and forget about it.
Yeah, you can split a terminal into panes. That's where this workflow came from. The difference is this is actual software now: real OS windows you can move between desktops, tile with one click, full-screen individually, and pick up with your mouse.
Open GG Coder on your side project in one window, your client's Next.js app in another, a Rust thing in a third, a landing page in a fourth. Each window runs its own agent, its own folder, its own model, its own history. Nothing bleeds between them.
Six projects, six different models (Claude, Codex, a local qwen3-coder, Gemini, Kimi, GLM), all going at the same time. And six isn't the ceiling either. Tile 2, 4, 6, or hit auto-arrange and it lays out however many you've got open.
The whole shell is Rust. No Electron, no bundled browser engine sitting in RAM per window. It uses the renderer your OS already ships, and each window's agent only costs you something while it's actually running. Six windows open is a normal Tuesday, not a fan event.
Not just GG Coder ones. It digs up everything you've touched in Claude Code and Codex too. Pick one, keep going.
Chat stays readable. Tools stream in a pinned panel at the bottom, so you see every file it touches and every command it runs without your conversation turning into a wall of JSON. Git branch, uncommitted count, open issues and PRs up top. Context %, thinking level and both models down bottom.
Anthropic, OpenAI/Codex, Gemini, Kimi, GLM, MiniMax, DeepSeek, Xiaomi MiMo, xAI, OpenRouter. OAuth or API key, your call. Swap models mid-conversation, nobody's stopping you.
Ollama, LM Studio, llama.cpp and vLLM get found automatically on their normal ports. No config, no flags. Real context windows and capabilities come from the server itself, so a model that can't call tools gets flagged right here instead of blowing up on your first prompt.
Flip Autopilot on and Ken (a mentor agent) reviews every finished run. If the work's not good enough he sends GG Coder straight back in with specific feedback, and it keeps going until he signs off. You go make coffee.
Real loop above: it built a rate limiter, Ken caught that the bucket was per-process and sent it back, it moved the thing into Redis and added a test, Ken signed off. Nobody typed anything in between.
Drag in a screenshot. Paste a design. Throw a video at it. Video goes straight to the models that handle it (Gemini 3.x, Kimi K3, MiniMax M3, MiMo-V2.5). For the ones that don't, the agent gets the file and reaches for ffmpeg itself.
Read-only poking around first, a written plan you approve, then it goes. For the stuff you really don't want it improvising on.
Every edit gets checked by a real language server. TypeScript ships in the box, zero setup, so type errors get caught and fixed in the same turn it created them. Python, Go, Rust and C/C++ kick in if their toolchain is on your PATH.
Paste any claude mcp add … line and it just works. Spawn subagents for parallel work.
Project memory, notes, chat export, prompt enhancement, your own slash commands in
.gg/commands/*.md.
Live usage meter in the title bar. How much of your 5-hour and weekly window you've burned, and when it resets. No more surprise rate limits.
XP, ranks and streaks for shipping. Sound. ASCII banners. Webcam gaze focus if you want your eyeballs to switch windows for you.
git clone https://github.com/KenKaiii/gg-framework.git
cd gg-framework
pnpm install
pnpm --filter @kenkaiiii/ggcoder build # build the sidecar first
cd gg-app && pnpm tauri devPackaging stuff (bundled Node runtime, single-file sidecar, code signing) is in gg-app/DISTRIBUTION.md.
Same agent, in your terminal. The app is the face, this is the engine.
npm i -g @kenkaiiii/ggcoder
ggcoderOAuth login so there's no API keys to paste, full terminal UI, tools, MCP, LSP diagnostics, session resume. → packages/ggcoder
Every layer ships on its own. Take one, take all of them.
| Package | What it does | README |
|---|---|---|
@kenkaiiii/gg-ai |
One streaming API for every provider up there | packages/gg-ai |
@kenkaiiii/gg-agent |
Agent loop with multi-turn tool execution | packages/gg-agent |
@kenkaiiii/gg-core |
Shared guts: model registry, OAuth, auth storage, paths | packages/gg-core |
@kenkaiiii/ggcoder |
The CLI, plus the sidecar the desktop app runs | packages/ggcoder |
@kenkaiiii/gg-boss |
Drives a bunch of workers across projects from one chat | packages/gg-boss |
@kenkaiiii/gg-ai (standalone)
└─► @kenkaiiii/gg-agent
└─► @kenkaiiii/gg-core
├─► @kenkaiiii/ggcoder ──► GG Coder desktop app ⭐
└─► @kenkaiiii/gg-boss
The desktop app forks zero agent logic. Windows, IPC and UI live in gg-app/.
Everything else is the exact same spine the CLI runs.
| You want to... | Use |
|---|---|
| Code with a real UI, across as many projects as you want | Download GG Coder |
| Code in your terminal | npm i -g @kenkaiiii/ggcoder |
| Run a bunch of agents across projects from one chat | npm i -g @kenkaiiii/gg-boss |
| Build your own agent that calls tools and loops | npm i @kenkaiiii/gg-agent |
| Stream from any LLM provider with one API | npm i @kenkaiiii/gg-ai |
pnpm install
pnpm build # tsc across all packages
pnpm check # typecheck
pnpm test # vitestTypeScript 5.9 · pnpm workspaces · Tauri 2 · React 19 · Vite 7 · Ink 6 · Vitest 4 · Zod v4
- YouTube @kenkaidoesai for tutorials and demos
- Skool community
MIT
Less bloat. More coding. Every model. Every project. One window each.
Rust under the hood, so it stays out of your way.
