OpenClaw Models & LLMs

Choose, Configure, and Optimize Your AI Stack

OpenClaw is model-agnostic: you can connect it to multiple model providers, switch models per workflow, and evolve your stack as new models appear.

This page is your practical guide to picking the right LLM for each task, setting up cloud or local inference, and keeping costs, latency, and safety under control as you scale from personal use to client delivery.

Why Your Model Choice Matters

In OpenClaw, the model is the “engine,” but your workflows, skills, and tool permissions determine what the agent can actually do.

Reliability

Fewer retries and better instruction following.

Cost Control

Lower token costs or free local inference.

Risk Management

Route sensitive data to private local models.

The OpenClaw Model Landscape

Typically easiest to start with: you bring an API key (or OAuth where supported) and get strong quality with minimal setup.

Best for: complex reasoning, long-form writing, multilingual quality, and “first try” reliability—especially when client deadlines matter.

Great when privacy, control, and predictable costs are the priority. OpenClaw supports local model setups through its gateway architecture.

Ollama Integration: One of the quickest paths to local inference. The integration can configure OpenClaw to use Ollama and start the gateway for you.

Recommendation: Use a large context window (at least 64k tokens) and models like qwen3-coder, glm-4.7, or gpt-oss variants.

How to Choose the “Right” Model

Most users get the best results by routing work by outcome, not by brand name. Use this decision rule:

Recommended Starting Configurations

You run the onboarding wizard, configure auth, and pick a default cloud model for most tasks.

Ideal when you want results today and don’t want to manage GPUs, drivers, or local runtime constraints.

You keep a premium cloud model as your “quality lane,” and add a local model for privacy-sensitive tasks and bulk processing.

This is where OpenClaw shines: mixing providers based on sensitivity, cost, and latency.

You run a strong local setup (often via Ollama), tune context length, and maintain one cloud model as a fallback.

Great for agencies and builders who want predictable unit economics and fewer vendor dependencies.

⚠️ Setup Pointers You Should Not Skip

  • Onboarding: Use openclaw onboard to set up model/auth, gateway settings, and workspace bootstrap.
  • Local Testing: Use openclaw dashboard to open the Control UI and chat locally at http://127.0.0.1:18789/.
  • Chat Channels: Node.js is recommended (Bun is not recommended due to known issues).
  • Ollama Launch: Use ollama launch openclaw (or alias ollama launch clawdbot) for a one-command local start.

Cost, Context & Reliability

If you want to make money with OpenClaw, your model strategy becomes your margin strategy.

Context Strategy: Large context windows increase token usage. Use them intentionally—reserve long-context models for tasks that truly need full-history reasoning. Split workflows: draft with a cheap model, polish with a premium one.

Reliability: Design skills with predictable inputs/outputs. Implement model failover early if uptime matters. Graceful degradation is better than single points of failure.