OpenClaw with LM Studio: Local Models, No Cloud, No API Costs

Running OpenClaw with LM Studio means zero API costs, full privacy, and complete offline operation. The tradeoff is hardware requirements and somewhat lower response quality compared to frontier cloud models.

If you've already set up OpenClaw with Ollama, LM Studio works the same way — it's the GUI alternative for the same local inference approach.

Why LM Studio Over Ollama?

Both tools run local models and expose an OpenAI-compatible API. The differences:

Choose LM Studio if you're on a desktop Windows or Mac and want a graphical way to manage models. Choose Ollama if you're deploying on a headless Linux VPS.

Step 1: Install LM Studio

Download LM Studio from lmstudio.ai. Versions available for macOS (Apple Silicon and Intel), Windows, and Linux.

Install and launch it. On first run, it will ask to download a model — you can skip this and do it in the next step.

Step 2: Download a Model

In LM Studio, click Search in the left sidebar. Browse available models. Good starting choices for OpenClaw use:

Search the model name, click the version with Q4_K_M quantisation (good quality-to-size ratio), and click Download.

Step 3: Start the Local Server

1. Click Local Server in the left sidebar (the <-> icon)
2. Select your downloaded model in the dropdown
3. Click Start Server
4. Note the server address — typically http://localhost:1234

The server exposes an OpenAI-compatible API. Leave LM Studio running.

Step 4: Configure OpenClaw

Add LM Studio as a provider in ~/.openclaw/config/providers.yml:

providers:
  lmstudio:
    api_key: "lm-studio"          # Any string — LM Studio doesn't validate API keys
    base_url: "http://localhost:1234/v1"
    default_model: "llama-3.1-8b-instruct"
    models:
      - id: "llama-3.1-8b-instruct"
        max_tokens: 4096

The api_key field can be any non-empty string — LM Studio's local server doesn't authenticate requests.

Set as active provider in config.yml:

llm:
  active_provider: "lmstudio"
  active_model: "llama-3.1-8b-instruct"

Restart OpenClaw and test.

Performance Expectations

Running inference locally is slower than cloud APIs. Typical generation speeds on consumer hardware:

A 200-token response at 40 tokens/sec takes ~5 seconds. That's acceptable for WhatsApp messages but noticeably slower than cloud APIs (typically 1–3 seconds).

CPU-only inference is usable for low-frequency tasks but too slow for a conversational AI agent.

Practical Tips

Keep LM Studio's server running: If you close LM Studio, OpenClaw loses its LLM. Either keep the window open or configure LM Studio to start the server on launch.

Model switching without restart: LM Studio lets you swap models without stopping the server. Change the model in the dropdown, wait for it to load, and OpenClaw's next request will use the new model automatically.

Use a dedicated SOUL.md preamble for local models: Local models are less reliable at following complex SOUL.md instruction sets than frontier models. Add a simplified version of your key rules at the start of each conversation context.

Hybrid approach: Use LM Studio/local models for routine, low-stakes tasks. Keep an OpenAI or Anthropic key configured as a fallback for complex requests:

llm:
  active_provider: "lmstudio"
  fallback_provider: "anthropic"   # Used when local model returns errors or low confidence

Related reading:

• Run OpenClaw with Ollama (Headless Local Models)
• Best LLM for OpenClaw: Full Comparison
• Managing OpenClaw Token Usage and Costs