Local Backends

Local backends run entirely on your machine. No API key required, no data leaves your system.

LM Studio

LM Studio is the recommended local backend for most users. Download models through its GUI, then GuardClaw connects to its built-in OpenAI-compatible server.

Setup

1. Download and install LM Studio
2. Download a model (recommended: qwen/qwen3-4b or qwen2.5-7b-instruct)
3. Go to Local Server tab → click Start Server
4. In GuardClaw:

guardclaw config set SAFEGUARD_BACKEND lmstudio
guardclaw config set LMSTUDIO_URL http://localhost:1234/v1
guardclaw config set LMSTUDIO_MODEL auto

Setting LMSTUDIO_MODEL=auto makes GuardClaw use whichever model is currently loaded in LM Studio.

Recommended Models

Model	Size	Accuracy	Speed
qwen/qwen3-4b-2507	4B	⭐⭐⭐⭐⭐	⚡⚡⚡⚡
qwen2.5-7b-instruct	7B	⭐⭐⭐⭐⭐	⚡⚡⚡
mistral-7b-instruct-v0.2	7B	⭐⭐⭐⭐⭐	⚡⚡⚡
llama-3.1-8b-instruct	8B	⭐⭐⭐⭐⭐	⚡⚡⚡
phi-3-mini-4k	3B	⭐⭐⭐⭐	⚡⚡⚡⚡

Avoid models under 3B — they struggle to produce reliable JSON output.

Using LM Studio as an OpenAI-compatible Endpoint

LM Studio can also serve as a proxy for any OpenAI-compatible API. Set LMSTUDIO_API_KEY if your endpoint requires auth:

guardclaw config set LMSTUDIO_API_KEY your-api-key

Troubleshooting

See the full LM Studio Troubleshooting guide.

---

Ollama

Ollama is ideal for Linux, Docker, and headless environments.

Setup

1. Install Ollama
2. Pull a model:

   ollama pull qwen2.5:7b

3. Configure GuardClaw:

guardclaw config set SAFEGUARD_BACKEND ollama
guardclaw config set OLLAMA_URL http://localhost:11434
guardclaw config set OLLAMA_MODEL qwen2.5:7b

Running Ollama

Ollama must be running before GuardClaw starts:

ollama serve   # starts the server

Or install as a system service — see Ollama docs.