pi-gmi-cloud

GMI Cloud provider package for pi.

Registers provider gmi against the OpenAI-compatible GMI Inference API and loads the live model catalog from GET /v1/models.

Install

pi install https://github.com/Xuanwo/pi-gmi-cloud

Pinned ref:

pi install git:github.com/Xuanwo/[email protected]

One-off without installing:

pi -e https://github.com/Xuanwo/pi-gmi-cloud

Auth

Create an API key in the GMI Cloud console, then either:

export GMI_API_KEY=your_key

or store it for pi:

/login gmi

pi also accepts a stored credential in ~/.pi/agent/auth.json:

{
  "gmi": {
    "type": "api_key",
    "key": "your_key"
  }
}

Usage

# List discovered models
pi --list-models gmi

# Start with a specific model
pi --provider gmi --model 'deepseek-ai/DeepSeek-V4-Pro'

Inside an interactive session:

/model

Select provider gmi, then any catalog model. Opening the model picker also refreshes the remote catalog when network access is allowed.

Behavior

| Item | Value | |------|--------| | Provider id | gmi | | API | openai-completions | | Base URL | https://api.gmi-serving.com/v1 | | Catalog source | GET /v1/models | | Cost units | Converted from GMI per-token USD to pi per-million-token USD | | Context window | From GMI context_length when present | | Auth | GMI_API_KEY, /login gmi, or auth.json |

Notes:

• Catalog fetch runs during extension load so pi --list-models gmi works immediately.
• Transient catalog failures keep the last successful in-memory list instead of crashing pi.
• Reasoning / image capability flags are inferred from model ids and names. Override later via models.json modelOverrides if needed.

Security

Pi packages run with full local permissions. Review this repository before installing on machines you care about.

License

MIT