I spent the last 3 months building my own AI agent harness, fully written in Go. It's called Forcefield, and I primarily built it because I had trouble using local AI models with agent harnesses like Claude Code. I found the configuration needed to get local models working frustrating, and I wanted something simpler.
The main thing I've been optimizing for is the runtime itself. It needs to have low overhead, fast startup, and a simple enough system that users can run it locally without an account or telemetry.
I'd be interested in feedback from people who use local AI coding tools. If you try Forcefield and find any bugs, please open an issue on GitHub. I'm also interested in feedback on the architecture and overall UX.
Comments URL: [https://news.ycombinator.com/item?id=49755908](https://news.ycombinator.com/item?id=49755908)
Points: 2
source & further reading
github.com — original article
Show HN: Skillmem – local memory for coding agents that stores how, not what
Show HN: Omabib, a Zotero alternative for Omarchy with agent support
RepoGuard – CLI to stop Cursor and Copilot from bypassing architecture