OpenClaw supports custom OpenAI-compatible model providers. LLMBase provides a dedicated agent endpoint for that use case, so an OpenClaw agent can use the current LLMBase agent model catalog with a separate, revocable chat-agent key.
This guide uses the LLMBase Agent API, which is different from the direct Inference API. It is intended for supported OpenAI-compatible agents using a Pro or Business Chat/Agent subscription. If an EU-hosted model is required for a task, choose an eligible model marked EU-hosted in the LLMBase model catalog and document that choice. Processing location remains model-specific.
1. Create the right LLMBase credential
In the LLMBase dashboard, open Agent Access and create a chat-agent key. It starts with:
llmbase_chat_...Use it only with this base URL:
https://llmbase.ai/api/v1/agentsThe key belongs to the agent API and uses the included Pro or Business Chat/Agent subscription budget. Keep it separate from a direct inference key, which starts with llmbase_... and belongs to https://api.llmbase.ai/v1.
This separation matters. A direct inference key and a chat-agent key have different endpoints and products. If OpenClaw reports an authentication error, checking the key prefix and base URL is the fastest first diagnosis.
2. Discover the current agent-compatible model list
Do not maintain an agent model allowlist from a blog post or a pasted example. Ask the Agent API for the current list before configuring OpenClaw:
curl https://llmbase.ai/api/v1/agents/models \
-H "Authorization: Bearer $LLMBASE_CHAT_AGENT_KEY"The response uses the OpenAI model-list shape and includes the models available to an agent key. Choose a model based on the work OpenClaw will do:
- Select a model that supports tools for agent loops that use tools.
- Check context length for long tasks and repository work.
- Check structured output or JSON support when the agent must hand data to another system.
- Select a model with the required input type before sending images or files.
- For EU-hosted work, choose an applicable model currently marked EU-hosted in the catalog.
If the model is not returned by the Agent API, it cannot be used with that chat-agent key.
3. Add LLMBase as an OpenClaw custom provider
OpenClaw’s Control UI and configuration format can evolve, but the connection values are stable. If you use the UI, add a custom provider with these fields:
| OpenClaw setting | Value |
|---|---|
| Provider name | LLMBase |
| API type | openai-completions |
| Base URL | https://llmbase.ai/api/v1/agents |
| API key | Your llmbase_chat_... key |
| Model ID | An ID returned by /api/v1/agents/models |
For file-based OpenClaw configuration, use the same values under models.providers. The model entry must use real metadata from the current Agent API instead of guessed limits:
{
models: {
mode: "merge",
providers: {
llmbase: {
baseUrl: "https://llmbase.ai/api/v1/agents",
apiKey: "llmbase_chat_...",
api: "openai-completions",
models: [
{
id: "<model-id-from-agent-models>",
name: "LLMBase agent model",
input: ["text"],
contextWindow: <context-length-from-agent-models>,
maxTokens: <max-output-length-from-agent-models>
}
]
}
}
}
}Use OpenClaw’s documented secret-reference or environment-substitution mechanism for apiKey when available. Do not commit a real agent key to a shared config repository. The snippet shows the field placement, not a safe place to store a credential.
4. Run a safe first agent task
Before attaching OpenClaw to a production workspace or giving it write-capable tools, verify the model connection with a read-only task:
Inspect the current project structure. List the three files most relevant to the requested change, explain why, and do not modify anything.
Confirm that OpenClaw selected the intended LLMBase model and returned a useful response. Then add one capability at a time. For example, first allow repository reading, then a test command that requires review, and only later consider any tool that changes data outside the agent.
The safest agent rollout is narrow and observable: the model is explicit, the allowed tools are minimal, and a person reviews any meaningful change.
5. Use EU-hosted model options responsibly
If an agent task has an EU-hosting requirement, make the model selection part of the task configuration rather than an informal preference. Record the exact model ID in the project documentation or OpenClaw allowlist, and review the current public model information whenever you change it.
For privacy, legal, or procurement decisions, use LLMBase’s current published materials and the organisation’s own approval process. A model badge helps make the selection visible; it does not replace an assessment of the specific work, data, and contractual requirements.
Troubleshooting OpenClaw with LLMBase
| Problem | What to check |
|---|---|
| Authentication fails | Use an llmbase_chat_... key with https://llmbase.ai/api/v1/agents, not an llmbase_... key or the direct API URL. |
| OpenClaw cannot find a model | Refresh the Agent API model list and add an exact returned ID to the custom provider configuration. |
| A capability is rejected | The model may not support the requested tool, JSON, image, or reasoning feature. Choose a model whose current metadata supports it. |
| The config needs model limits | Copy context and output limits from the Agent API response instead of estimating them. |
| You need an EU-hosted model | Use a model currently marked EU-hosted in the LLMBase catalog, then keep that ID explicit in the OpenClaw configuration. |
FAQ: OpenClaw and LLMBase
Can OpenClaw use LLMBase models?
Yes. Configure LLMBase as an OpenAI-compatible provider using the Agent API base URL and a chat-agent key.
Do I use the normal LLMBase API key in OpenClaw?
Not for the subscription-backed agent setup in this guide. Use an llmbase_chat_... chat-agent key at https://llmbase.ai/api/v1/agents.
Why should I load models from the Agent API?
The Agent API exposes the current models available to chat-agent keys. It prevents a configuration from using a model that is not eligible for that endpoint.
For the full LLMBase contract, see Agent integrations, agent setup examples, and models for agents.
