Install

pip install -U localcode

Machine	Mac with Apple Silicon
Unified memory	At least 16 GB
Python	3.10 or newer
Disk	Space for one model - the smallest recommended GGUF is about 7.4 GB

Apple Silicon is the supported platform. Metal-accelerated inference works only on Mac. localcode also installs and runs on Linux in CI for development, but Linux is not the product platform.

Open your repo

cd ~/work/some-project
localcode

Choose a project whose tests already pass. localcode uses your repo’s own checks as proof.

Choose a model

The model picker opens on first launch. localcode checks your Mac’s unified memory and marks a recommended model with a star. Use the arrow keys to choose one, then press Enter.

localcode downloads the model’s GGUF from Hugging Face - the only step that needs the network, and only once per model. Then it starts the included llama-server at http://localhost:8081 and connects the agent to it.

Learn more in Models.

Start building

Enter your request in the chat screen. Include the file name and the check to run.

> Implement the retry decorator in retry.py so every test in test_retry.py
  passes. Do not modify test_retry.py. Then run: pytest -q

Watch it verify

The model reads the stub and tests. It writes the code, runs pytest -q, and reports what it checked.

Qwen3.6-35B-A3B (IQ2_M) runs locally on 127.0.0.1:8081. It uses four tool calls, takes 11.5 s, and uses 276 tokens. The repository’s tests fail before the turn and pass after it.

Key commands

Command	What it does
/status	Shows server health, the current model, and performance settings
/model	Lists models or switches models (/model qwen)
/permissions	Turns command approvals on or off
/clear	Clears the conversation history
/exit	Quits

See the full list: Slash Commands.

Next

• Models - find the best model that fits on your Mac.
• Network Boundary - learn what leaves your machine and when.