How to Setup OpenClaw using Custom Provider

Do not index
If your model provider isn’t listed in OpenClaw by default, you can still connect it using a custom provider. This is useful for services that expose OpenAI-compatible or Anthropic-compatible APIs.
Here’s how to set it up on Agent37.
Follow the steps below or watch the video walkthrough:

Video Walkthrough

Video preview

Step 1: Open your Agent 37 instance

  1. Go to the Agent37 dashboard and sign in.
  1. Open an existing instance, or create a new one.
  1. Launch Terminal inside the instance.

Step 2: Configure OpenClaw

In Terminal, run:
openclaw configure
When prompted, choose the following:
  • Where will the gateway run?Local
Continue with:
  • Select Section to Configure?Model
  • Model / Auth ProviderCustom provider
  • API Base URL → your provider’s base API URL, such as:
According to OpenClaw’s onboarding flow, Custom provider is the correct choice for endpoints that are not listed directly, including providers that expose OpenAI-compatible or Anthropic-compatible APIs.
Next, complete the remaining prompts:
  • How do you want to provide this API key now?Paste API key now
  • Endpoint-compatibility
    • Choose OpenAI-compatible if your provider exposes an OpenAI-style /v1 API
    • Choose Anthropic-compatible if it uses Anthropic’s Messages API
    • Choose Unknown if you want OpenClaw to auto-detect compatibility
  • Enter API key → paste the API key you copied from your provider
  • Model ID → enter the exact upstream model name
    • Example: Minimax-M2.7
  • Alias (optional) → enter a friendly display name for the model
  • Endpoint ID → enter a short provider identifier such as: myproxy or minimax

Step 3: Verify the model

Run:
openclaw models list
Your custom model should appear as:
<endpoint-id>/<model-id>
Example:
minimax/Minimax-M2.7
If you want to make it your default model, run:
openclaw models set minimax/Minimax-M2.7

Step 4: Test the setup.

Run:
openclaw dashboard
Then open the dashboard and send a test prompt to confirm everything is working correctly.