# OpenClaw

Add Vultr Serverless Inference to OpenClaw from ClawHub: the live model list with context windows, prices, input types and thinking levels.

`@vultr/model-provider-openclaw` is an OpenClaw provider plugin. It needs OpenClaw 2026.9.4 or later.

## Set it up

```bash
openclaw plugins install clawhub:@vultr/model-provider-openclaw --accept-capabilities
export VULTR_INFERENCE_API_KEY=...
openclaw models list --all --provider vultr --refresh
openclaw infer model run --local --model vultr/glm-5.3 --thinking high --prompt "hello"
```

`openclaw onboard --vultr-api-key <key>` stores the key instead of the environment variable. A running gateway picks up the plugin on its next start. To install from GitHub rather than ClawHub, `--force` is required:

```bash
openclaw plugins install git:github.com/vultr/model-provider-openclaw --force --accept-capabilities
```

## Pick a model

Models are `vultr/<id>`. `--thinking` takes the model's reasoning levels, and `off` where the model can answer without thinking.

## What it maps

| OpenClaw | From the catalog |
| --- | --- |
| Context window | The model's |
| Output limit | The smallest of the model's output limit, 65536, and a quarter of the context window |
| Cost | Prices per million tokens, including cached input |
| Input | Text and images, as the model accepts them |
| Thinking levels | Reasoning levels |

The output limit is lower than the model's own because OpenClaw sends it as `max_tokens` without leaving room for the prompt; for most models the published limit is the whole context window. The last good list is kept in the agent directory, under `cache/vultr-model-catalog.json`.