Why I built a local-first AI agent instead of another cloud one

Most AI agents run in the cloud. Your prompts, your files, your context and all sent off to a provider, and you have to trust a black box with your data and your machine!

Our team wanted the opposite. So we've been building Atomic Agent: an open-source (MIT) AI agent runtime that runs locally, on your own hardware, on top of llama.cpp. It drives your desktop: browser automation, file editing, shell commands; with the risky actions gated behind an approval step, and it keeps sessions, memory, and traces on your own disk. It works with local models, and also with OpenAI-compatible / OpenRouter endpoints if you want them. CLI and TUI, macOS and Linux, with a Windows developer preview.

The bet behind it: running an agent shouldn't mean giving up control of your data, or trusting something you can't read. Everything is open source, so you can inspect exactly what it does before it does it.

Repo: https://github.com/AtomicBot-ai/atomic-agent

Site: https://atomicagent.io

It's early and there's a lot still rough. Building in public here and would genuinely love feedback from this community.