VibeCoder: Local LLM Coding Assistant

The Problem

I wanted an AI coding assistant that never sent my code to anyone else's servers. So I built VibeCoder to run entirely on local LLMs.

Claude and Cursor are better at the actual code generation, no argument there, but every query runs through their infrastructure. For proprietary work or anything sensitive, that is a real constraint. VibeCoder runs on Ollama instead: no cloud, no API keys, no data retention policies to read.

How It Works

VibeCoder is a Python terminal app that routes input by prefix. You type edit:, run:, create:, search:, or plan: and it dispatches to the right handler.

The edit: command reads a file, sends it to the local model with your instruction, and shows a colored diff before writing anything. Line-range syntax like [main.py:10-20] points the model at specific code instead of dumping a whole file into a small context window. The plan: command breaks a request into a JSON array of steps and executes each one with approval before anything destructive. Command execution goes through a whitelist and blocks dangerous calls like rm, and every write requires explicit confirmation.

The Honest Tradeoff

Local execution keeps your code on your machine, but you pay for it in quality. A 7B model on a laptop does not match a frontier model behind an API.

Local models need real hardware, so expect 16GB+ of RAM for anything useful. VibeCoder works best when your machine can comfortably run at least a 13B model. If privacy is not a hard requirement, a cloud tool will write better code, and that is the honest call.

What I'd Do Differently

Context management is the weak spot. It just sends recent messages to the model with no summarization, so long sessions degrade as the window fills up.

I would rebuild that first, then add a test: command for unit tests and a commit: command for Git, both obvious gaps. I wrote up the full architecture in the build post, and if you want the same local-first philosophy applied to project management, take a look at AGINEAR.