KenKaiii/gg-framework

What it is

GG Coder is a desktop coding agent that works with the AI subscriptions or API keys you already have, including Claude, ChatGPT, Gemini, Grok, Kimi and local models. The repository also publishes TypeScript packages for building agents without the desktop app: @kenkaiiii/gg-ai for provider connections and streaming, @kenkaiiii/gg-agent for tool-using agents, @kenkaiiii/gg-core for model selection and authentication, and @kenkaiiii/ggcoder for the coding agent, sessions and tools (20).

Beyond code, the app bundles GG Goals for handing over larger objectives to worker agents, a Chat mode, and GG Motion for turning ideas into videos. Two features are emphasized as differentiators: Agent Steroids, which searches open-source code locally without an MCP server or rate limit, and Ken Autopilot, which continues work until results are verified.

Quick start

Install the desktop app from the latest release: open the .dmg on Apple Silicon macOS, run the .exe installer on Windows, or install the .deb/.rpm on Linux x86_64 (16)(4). Then open the app, connect an AI provider, and choose Code, Chat or Motion (19).

To run from source, install Node.js, pnpm and the Tauri prerequisites, then:

git clone https://github.com/KenKaiii/gg-framework.git
cd gg-framework
pnpm install
pnpm build
pnpm --filter gg-app tauri dev

To embed the engine in another project, start with npm install @kenkaiiii/gg-ai @kenkaiiii/gg-agent; the package READMEs cover setup and usage (25)(21)(22).

Constraints

• macOS builds target Apple Silicon, and Linux builds target x86_64.
• Building from source requires the Tauri toolchain in addition to Node.js and pnpm (17).
• The desktop app and framework use AI providers under your own accounts, so provider access is a prerequisite.
• The project is licensed under AGPL-3.0: changes you distribute or run as a network service must be shared under the same license (20).
• The benchmark claims (fastest and leanest of 20 agents, 102 s and $0.014 per suite on Claude Haiku 5.5) come from the maintainer's own single race on 4 November 2026 with 7 tasks and 5 runs each; they are not independent measurements (14).

HotRepo's take

On 5 November 2026 the repository reached 6th place on the TypeScript daily list with 27 stars gained that day, out of 191 total. That is a single snapshot rather than a sustained trend, and the full-language daily rank is not recorded. The benchmark table is unusual in its specificity—20 named tools, per-run cost and timing, and a stated model and task set—which makes the performance claims checkable rather than vague, even though the race was run by the maintainer. The split between a desktop product and independently installable packages means readers can evaluate the engine without adopting the app.