Built-in MLX Backend

The built-in backend downloads and runs an LLM directly inside GuardClaw — no LM Studio or Ollama needed. Uses Apple Silicon MLX for fast, efficient inference.

Requires: macOS with Apple Silicon (M1/M2/M3/M4).

Setup

The setup wizard can configure this automatically. Or manually:

guardclaw config set SAFEGUARD_BACKEND built-in

On first use, GuardClaw will prompt you to download a model. Models are stored at ~/.guardclaw/models/.

Model Management

guardclaw config llm   # interactive model picker (download, load, unload)

Or via the dashboard Settings → Model panel:

• Browse available models
• Download with progress bar
• Load / unload with one click

How It Works

GuardClaw spawns mlx_lm.server on port 8081 using a Python venv at ~/.guardclaw/venv/. The venv is created automatically on first use.

The model file (~2–4 GB) is downloaded once and cached at ~/.guardclaw/models/.

Disk Usage

Path	Contents
~/.guardclaw/venv/	Python venv + mlx_lm (~500 MB)
~/.guardclaw/models/	Downloaded model weights (2–4 GB each)

To free space:

guardclaw config llm   # → unload / delete model