# pi

Use Vultr Serverless Inference in the pi coding agent: one package install, and the live model list with context windows, prices and thinking levels.

`@vultr/model-provider-pi` is a pi package that registers the `vultr` provider, from the live model list on every start.

## Set it up

```bash
pi install npm:@vultr/model-provider-pi
export VULTR_INFERENCE_API_KEY=...
pi --list-models vultr
pi --model vultr/glm-5.3:high
```

pi lists only providers it has a key for, so `--list-models` shows no `vultr` models until `VULTR_INFERENCE_API_KEY` is set. To install from GitHub instead: `pi install git:github.com/vultr/model-provider-pi`.

## Pick a model

Models are `vultr/<id>`, and a thinking level follows a colon: `vultr/glm-5.3:high`. pi shows only the levels the model accepts, and `off` where the model can answer without thinking.

## What it maps

| pi | From the catalog |
| --- | --- |
| Context window and max tokens | The model's. pi leaves room for the prompt itself |
| Cost | Prices per million tokens, including cached input |
| Input | Text, and images on models that read them |
| Thinking levels | Reasoning levels |

The last good list is kept in pi's agent directory, under `cache/vultr-model-catalog.json`. With neither the network nor that file, pi starts without the provider and says so in one line.