Local Coder
A desktop coding agent that runs entirely against local models. No API keys, no per-token bill, no code leaving the machine.
Local Coder is a desktop coding agent that talks to models running on your own hardware.
Why build it
Two reasons, and only one of them is cost.
The obvious one is that a coding agent billed per token gets expensive on long sessions. The better one is that some code should not leave the machine it lives on. If the model is local, that question does not arise, and you do not have to trust a policy document about it.
How it works
An Electron shell around a local server that speaks to LM Studio, with the agent loop, tool calling and file editing handled locally. It ships as a portable executable, so there is no install step and nothing to uninstall.
Alongside the core agent it grew a few things I needed: an image generation integration, model and adapter fetching, batch generation scripts, a model-swapping configuration generator, and a mobile mode for driving it from a phone on the same network.
The honest assessment
A locally hosted model on consumer hardware is not competitive with a frontier model on hard reasoning. It is genuinely good at the bulk of day-to-day work: mechanical refactors, writing tests, moving code between files, explaining an unfamiliar function.
The value is knowing where the line is and routing work accordingly, rather than pretending the local model is something it is not.