BotBytes Coder — An Offline Coding Agent by Muhammad Hasnat AkhtarBotBytes Coder — An Offline Coding Agent by Muhammad Hasnat Akhtar

BotBytes Coder — An Offline Coding Agent

Muhammad Hasnat Akhtar

Muhammad Hasnat Akhtar

Every hosted coding assistant has the same problem for anyone working under a confidentiality agreement: to help you, it has to read your code, and reading it means sending it somewhere. For a lot of real work that ends the conversation.

What I built here

A coding agent that runs on the machine it edits. The model is local, there is no API key, and nothing is sent anywhere. It has a command-line mode and a browser UI, and a single diagnostic command that checks the install and prints the fix next to each problem it finds rather than a stack trace.

The guarantees, and how each one is actually held

A promise in a README is a hope. Each of these is held by code, and each has its own test:
The model never writes a file on its own. Its answer comes back as a diff in the UI, and the write happens on approve. Not a setting that defaults to safe — there is no always-yes path, and the hidden environment variable that would re-enable one is closed too.
Protected paths cannot be read. A pre-execution hook inspects every tool call, including the full text of a shell command, and refuses anything that resolves into a protected directory. The guard is not a prompt the model is asked to respect; it runs before the tool does.
Nothing reaches the internet. Update checks, analytics and price lookups are all switched off.
Work is not lost. Before a run, it checks for uncommitted changes and asks first.

Fine-tuning, kept private

The agent can be taught the owner's own coding rules. The training material is prepared and the job is run on external GPU, and the design constraint there is the same one as everywhere else: the assistant orchestrating the run is never allowed to read the training examples themselves. It sees counts, checksums and a pass or fail — a problem is reported as a position in the data, not as its contents. The dataset and the job are both private, and that is verified after upload rather than assumed.

Why I built it

I wanted an agent I could point at work that is not allowed to leave the building. Once that is the requirement, "local model" is the easy part — the hard part is proving the thing cannot phone home, cannot read what it was told not to read, and cannot write a file behind your back. That is what most of this build actually is.
This is my own product, not client work.
Like this project

Posted Oct 15, 2026

My own product: a coding agent that runs entirely on the machine it edits — no cloud call, no API key, and no file written until a human approves the change.