Every "local" AI coding tool I tried stopped at serving a model — Ollama, LM Studio — with no real agent loop on top. And every full agent I tried sent my code to someone else's cloud whether I wanted it to or not. I wanted both: a real agent that reads, acts, and checks its own work, and the option to run it entirely on my own machine. What problem were you solving? Trust. Not "is the model good," but "do I actually know what it's about to do to my files." Most tools either ask you to blindly approve everything or don't ask at all. I wanted a permission engine that's plain, readable code — four modes, a command-risk classifier — not an LLM's guess about what's safe, plus a real colored diff before anything gets written, and a checkpoint I can revert even if the task created new files. How did the approach evolve? The permission engine and diff review came first and felt sufficient — until I actually dogfooded it on a real project and watched a smaller local model describe a file change in prose without calling the edit tool at all. That's what pushed me to add a corrective nudge when a model claims it edited something but didn't, and eventually an auto-verify step: after a successful edit, localagent runs the project's own test command automatically and feeds the real result back before the model is allowed to say it's done. Small local models are genuinely less reliable on complex changes — I'd rather say that outright in the UI than let someone find out the hard way. Still early — beta, unsigned builds today — so I'd love feedback, especially on where the permission model should go next.

localagent - A local first coding agent bring your own model no cloud

Replies