> ## Documentation Index
> Fetch the complete documentation index at: https://docs.tokenlab.sh/llms.txt
> Use this file to discover all available pages before exploring further.
# TokenLab Provider Setup
> Configure TokenLab as a custom provider in OpenClaw using OpenAI, Responses, Anthropic, and Gemini-compatible routes
This guide is for **self-hosted OpenClaw** users who want to connect TokenLab as their AI provider.
## Recommended: install the plugin
For standard OpenClaw agent execution, install the TokenLab provider plugin:
```bash theme={null}
openclaw plugins install @tokenlabai/openclaw-provider
openclaw onboard --tokenlab-api-key "$TOKENLAB_API_KEY"
```
The plugin discovers TokenLab's live chat-model catalog and keeps 18 fallback models for offline discovery. Models use the `tokenlab/` format.
[npm package](https://www.npmjs.com/package/@tokenlabai/openclaw-provider) · [ClawHub listing](https://clawhub.ai/tokenlabai/plugins/openclaw-provider) · [Source code](https://github.com/hedging8563/tokenlab-openclaw-provider)
## Manual provider configuration
Use manual `models.providers` entries when you explicitly need separate **Responses API**, **Claude native**, **Gemini native**, or **MiniMax native** routes.
If you choose manual configuration, configuring `tokenlab` alone is enough. Add the other providers only when you explicitly need **Responses API**, **Claude native**, **Gemini native**, or **MiniMax native** behavior.
| Provider | OpenClaw `api` | Best for | baseUrl |
| -------------------- | ---------------------- | ---------------------------------------------------------------- | ---------------------------- |
| `tokenlab` | `openai-completions` | GPT, DeepSeek, Qwen, and most OpenAI-compatible calls | `https://api.tokenlab.sh/v1` |
| `tokenlab-responses` | `openai-responses` | OpenAI Responses workflows that expect `/v1/responses` semantics | `https://api.tokenlab.sh/v1` |
| `tokenlab-claude` | `anthropic-messages` | Native Claude Messages API | `https://api.tokenlab.sh` |
| `tokenlab-gemini` | `google-generative-ai` | Native Gemini API format | `https://api.tokenlab.sh` |
| `tokenlab-minimax` | `anthropic-messages` | Native MiniMax routing | `https://api.tokenlab.sh` |
Use the `/v1` suffix only for `openai-completions` and `openai-responses`.
Native providers such as `anthropic-messages` and `google-generative-ai` should use `https://api.tokenlab.sh` without `/v1`, otherwise OpenClaw may construct the wrong provider path.
## Prerequisites
* A self-hosted OpenClaw instance
* A TokenLab API Key — [Get one here](https://tokenlab.sh/dashboard)
## Configuration
Edit your OpenClaw config:
* **Self-hosted**: `~/.openclaw/openclaw.json`
Add TokenLab providers under `models.providers`:
```json5 theme={null}
{
agents: {
defaults: {
model: {
primary: "tokenlab-claude/claude-sonnet-4-6"
}
}
},
models: {
mode: "merge",
providers: {
tokenlab: {
api: "openai-completions",
baseUrl: "https://api.tokenlab.sh/v1",
apiKey: "sk-your-tokenlab-key",
models: [
{ id: "gpt-4o", name: "GPT-4o" },
{ id: "deepseek-r1", name: "DeepSeek R1" },
{ id: "qwen3-coder-flash", name: "Qwen 3 Coder Flash" }
]
},
"tokenlab-responses": {
api: "openai-responses",
baseUrl: "https://api.tokenlab.sh/v1",
apiKey: "sk-your-tokenlab-key",
models: [
{ id: "gpt-4o", name: "GPT-4o (Responses)" },
{ id: "gpt-5.2", name: "GPT-5.2 (Responses)" }
]
},
"tokenlab-claude": {
api: "anthropic-messages",
baseUrl: "https://api.tokenlab.sh",
apiKey: "sk-your-tokenlab-key",
models: [
{ id: "claude-sonnet-4-6", name: "Claude Sonnet 4.6" },
{ id: "claude-opus-4-6", name: "Claude Opus 4.6" }
]
},
"tokenlab-gemini": {
api: "google-generative-ai",
baseUrl: "https://api.tokenlab.sh",
apiKey: "sk-your-tokenlab-key",
models: [
{ id: "gemini-3.5-flash", name: "Gemini 3.5 Flash" },
{ id: "gemini-2.5-pro", name: "Gemini 2.5 Pro" }
]
},
"tokenlab-minimax": {
api: "anthropic-messages",
baseUrl: "https://api.tokenlab.sh",
apiKey: "sk-your-tokenlab-key",
models: [
{ id: "minimax-m3", name: "MiniMax M3" }
]
}
}
}
}
```
All 5 providers use the **same API Key**. You only need one TokenLab account.
The `models` arrays above only show common examples. Add more model IDs to each provider as needed.
## Using Models
OpenClaw still references models with the `provider/model` format:
```json5 theme={null}
{
agents: {
defaults: {
model: {
primary: "tokenlab-gemini/gemini-3.5-flash"
}
}
}
}
```
### Model Examples
| Provider | Model reference | Description |
| -------------------- | ----------------------------------- | ---------------------------- |
| `tokenlab` | `tokenlab/gpt-4o` | OpenAI-compatible route |
| `tokenlab` | `tokenlab/deepseek-r1` | DeepSeek reasoning model |
| `tokenlab-responses` | `tokenlab-responses/gpt-4o` | Responses API route |
| `tokenlab-claude` | `tokenlab-claude/claude-sonnet-4-6` | Native Claude Messages route |
| `tokenlab-gemini` | `tokenlab-gemini/gemini-3.5-flash` | Native Gemini route |
| `tokenlab-minimax` | `tokenlab-minimax/minimax-m3` | Native MiniMax route |
Browse all available models at [tokenlab.sh/models](https://tokenlab.sh/en/models).
## When to Use Which Provider
* `tokenlab`: default choice for most general-purpose agent and chat use cases.
* `tokenlab-responses`: use when your OpenClaw workflow explicitly depends on OpenAI Responses semantics.
* `tokenlab-claude`: use when you want Claude's native Messages behavior.
* `tokenlab-gemini`: use when you want Gemini-native request/response formatting or existing Gemini-style integrations.
* `tokenlab-minimax`: use when you want MiniMax on its native route.
If you do not need Gemini-native behavior, you can still call Gemini models through `tokenlab/gemini-*` on the OpenAI-compatible route.
## Common Mistakes
Current OpenClaw docs use `models.providers`. If you keep the older top-level `providers` array format, OpenClaw may ignore the config or fail to resolve the provider prefixes as expected.
`openai-responses` maps to TokenLab's `/v1/responses` path, so `tokenlab-responses` must use `https://api.tokenlab.sh/v1`.
`anthropic-messages` and `google-generative-ai` should use `https://api.tokenlab.sh` without `/v1`. Adding `/v1` can produce incorrect request paths.
Yes. Current OpenClaw documentation still includes the built-in `google` provider and also supports custom providers using `api: "google-generative-ai"`. So `tokenlab-gemini` remains a valid native Gemini route for OpenClaw users.
## Verify Setup
After saving the config, restart your OpenClaw instance and test with a simple message. If you see a response, the provider is configured correctly.
```bash theme={null}
# Self-hosted: restart the service
systemctl --user restart openclaw # Linux
launchctl stop cc.tokenlab.openclaw && launchctl start cc.tokenlab.openclaw # macOS
```
## Next Steps
Once OpenClaw is connected, these guides help you use TokenLab more effectively:
* [API Formats](/guides/api-formats) — understand the differences between OpenAI, Responses, Anthropic, and Gemini routes
* [IDE / SDK Compatibility](/guides/ide-sdk-compatibility) — see when `/v1/responses` is the better fit
* [Error Handling](/guides/error-handling) — learn common failure modes and recovery patterns
* [Models Overview](https://tokenlab.sh/en/models) — browse model IDs before wiring them into agents