pi-kit: A Local-First Coding Agent UI

Over the last couple of months, we worked extensively with local models under Windows. There, we used our friends GLM, Qwen, and DeepSeek, which work well for our use case. The main issue was the Windows command prompt, which isn't great compared with Linux or Mac. Our harness was/is https://pi.dev, which we love because it adds little overhead and feels fast; plus, it's made in Austria—specifically, Styria, where I also live—so it was a no-brainer.

After the pain with Windows and the command prompt, we decided to build a small UI over Pi so I could enjoy my life again. From a technical standpoint, I wanted to use tech and open-source projects I used and liked before, like Angular, NgRx (Rainer is also a maintainer there), NestJS, and CopilotKit Angular (which we also maintain together with CopilotKit). But for communication, we decided to use the AG-UI protocol, which is a great standard. During the build process, the Pi project released Pi Durable (https://earendil.com/posts/pi-durable/), so it was a no-brainer to use it for our optional subagents, which we'd call robust and durable.

What's inside pi-kit

The most important thing for us was supporting local models, since we also run workshops in that area; we also support llama.cpp. But installation and configuration can be cumbersome, so we added an option to download Hugging Face models directly in pi-kit, along with their configuration.

Since we are still doing project work for customers and developing web apps for that purpose, we wanted pi-kit to have eyes on what we are building. This is our artifact browser, which supports screenshots and console-log retrieval.

Code review is very important, and we still stand by it; that means we support a diff viewer where we can batch the comments during a review to improve the workflow. We also have plans for the future, including Sheriff (Rainer created it to protect our architecture).

But we could go on and on; here is a list: Background agents, Computer use, Remote Environments (iPhone / Android), MCP, A2UI, Gen-UI, just to name a few.

We built a whole webpage at https://pi-kit.dev/ with pi-kit; we also generated the videos with qwen-tts. Go check it out. Please use it and give us feedback. Pi-kit works really well with local models, as well as your ChatGPT subscription or Pi Radius, a commercial service from Pi (https://earendil.com), so we can support them.

Special thanks to Markus Ecker from CopilotKit for helping with Linux use and inspiring many of the feature ideas.

Everything is open source, MIT-licensed; take it, fork it, contribute back.