Qodey

A coding agent for your terminal that runs on your own models. It uses Ollama, so everything stays on your machine — no accounts, no API keys, no cloud.

You type qodey, describe what you want, and it pokes around your project, edits files, runs commands, and fixes its own errors until the job is done (or it gets genuinely stuck, in which case it says so).

You need

• Python 3.9+
• Ollama installed and running (ollama serve)
• At least one model, e.g. ollama pull qwen2.5-coder:32b

One honest note: coder models work noticeably better than plain chat models. A small chat model can still use Qodey, it will just fumble the tools more often.

Install

pip install qodey
qodey

From source, if you want to hack on it:

pip install -e .
python -m qodey

Or isolated with pipx:

pipx install qodey

What it does

• Figures out what's in your folder (React app, Python script, Rust crate, empty dir, …)
• Reads, writes, and edits files; creates folders; searches code
• Runs shell commands — installs, builds, tests — and actually reads the output when they fail
• Asks before anything destructive. Reads never ask; writes and commands do, unless you switch to auto mode

A typical exchange:

❯ add validation to the signup form

  ◉ Reading src/Signup.tsx…
  ✓ Reading src/Signup.tsx
  ◉ Editing src/Signup.tsx…
  ✓ Editing src/Signup.tsx

Done — email and password are validated on submit, errors show
under each field. `npm test` passes.

Commands

Permissions

Three modes, default is confirm:

• confirm — reads happen freely, writes and commands ask first
• auto — everything runs without asking (your funeral with rm -rf)
• safe — asks about practically everything

Change with /permissions.

Config

Lives at ~/.qodey/config.json. You usually don't need to touch it — /settings covers the common stuff — but it's plain JSON if you want to:

{
  "ollama_url": "http://127.0.0.1:11434",
  "default_model": "qwen2.5-coder:32b",
  "temperature": 0.2,
  "permission_mode": "confirm",
  "plan_mode": true,
  "max_agent_steps": 25
}

How it works

Each turn, Qodey sends the model a short description of your project (file list, detected stack — never whole files), the model answers with text plus tool calls, Qodey runs them, feeds the results back, and repeats until there's nothing left to do. If a tool call gets denied or a command fails, the model sees that and adjusts.

License

MIT