How to Configure the Right Model in OpenClaw for Reliable Agent Performance

After installing OpenClaw, many users open the configuration page and are confronted with a long list of model names—Claude, GPT, DeepSeek, Gemini, Qwen, etc.—without knowing which one to choose. Selecting the wrong model often leads to slow responses, tasks aborting midway, or unexpected behavior, which is not a deployment issue but a model issue.

OpenClaw itself is merely a framework that receives commands, decomposes tasks, calls tools, and executes actions; the actual reasoning depends entirely on the AI model you configure. Think of OpenClaw as a car chassis and the model as the engine—an underpowered engine cannot make the car fast or climb hills.

Choosing an inappropriate model can cause:

Misinterpretation of instructions, causing the task to deviate.

Loss of context in long conversations.

Injection of malicious commands from web content.

Tool‑call failures, such as sending unintended emails.

Because OpenClaw’s agent loop runs continuously, a single judgment error propagates to all subsequent actions.

The entry point for model configuration is the openclaw models command family. The following workflow, illustrated with the Qwen provider, also applies to other supported providers like OpenAI, Anthropic, and Gemini.

1. List supported model providers

Run: openclaw models list-providers Sample output:

qwen-portal Qwen (Tongyi Qianwen) — requires OAuth
openai OpenAI GPT series
anthropic Claude series
google Gemini series
local Local LLMs (via Ollama, LM Studio, etc.)

Tip: Use openclaw plugins list to check if additional model adapters are installed.

2. Authenticate with a provider (example: qwen-portal)

openclaw models auth login --provider qwen-portal

This opens a browser or prompts for an Access Token / Refresh Token. Qwen requires OAuth; some versions may also accept an API‑Key depending on the acp configuration.

If you see “refresh token expired”, the previous token is invalid and you must log in again.

3. Verify authentication status

openclaw models auth status --provider qwen-portal

Successful output should resemble:

Provider: qwen-portal
Status: authenticated
Expires: 2026-03-20T08:00:00Z

4. Set the default model (optional)

Edit ~/.openclaw/config.yaml or run one of the commands below:

openclaw config set model.default qwen-portal/coder-model

or

openclaw config set model.vision qwen-portal/vision-model

You can also specify the model for a single run:

openclaw agent run --model qwen-portal/coder-model "write a Python quicksort"

5. Advanced: Custom model routing / proxy

To route a private model through a local proxy (e.g., localhost:8000/v1), create or edit ~/.openclaw/models/qwen-portal.yaml:

endpoint: http://localhost:8000/v1
api_key: sk-xxx
headers:
  Authorization: Bearer sk-xxx

After saving, restart the OpenClaw Gateway or reload the configuration for the changes to take effect.