@anandamw/custom-provider

A Pi extension for managing providers and models through a unified /rakit command. Supports OpenAI-compatible local servers (Ollama, llama.cpp, vLLM), hosted gateways, and Pi-native models.json providers.

Install

pi install npm:@anandamw/custom-provider

Or select Custom Provider in the Pi Rakit installer.

Command: /rakit

Type /rakit inside Pi to open the management menu:

Select provider or manage
> Claude (Anthropic) — 5 models
  ChatGPT (OpenAI) — 12 models
  My Local Server — 2 models | http://localhost:11434/v1 (custom)
  Add custom provider
  Manage custom providers
  Manage models.json
  Exit

Selecting a provider

Pick any provider from the list to see its models with detail:

> ag-claude — ctx:68k max:16k
  gemini — ctx:128k max:8k | 🧠

Select a model to switch the active session to that provider/model.

Add custom provider

Prompts for:

1. Provider ID
2. Provider name
3. API URL (http/https only)
4. API key
5. Model — auto-discovered from GET <baseUrl>/models, or enter manually

After saving, the provider is registered and the model is selected for the current session. Saved to ~/.pi/agent/settings.json.

Manage custom providers

Add, edit, or delete custom providers and their models. Each provider supports multiple models. Editing a model re-prompts all fields with current values as defaults.

Manage models.json

Manage Pi-native providers defined in ~/.pi/agent/models.json.

9routerclaude — 1 model | http://localhost:20128/v1

Inside a provider, select a model to Edit or Delete it, or choose + Add new model.

When adding or editing a model, fields are:

| Field | Description | | --- | --- | | Model ID | Unique identifier | | Model name | Display name | | Context window | Positive integer (e.g. 68000) | | Max tokens | Positive integer (e.g. 16384) | | Reasoning | true or false | | Input type | text or text + image |

Changes to models.json take effect after reloading Pi.

Environment variables (default provider)

The extension also registers a default OpenAI-compatible provider via environment variables:

| Variable | Default | Description | | --- | --- | --- | | PI_RAKIT_PROVIDER_ID | rakit-openai | Provider identifier | | PI_RAKIT_PROVIDER_NAME | Pi Rakit OpenAI Compatible | Display name | | PI_RAKIT_PROVIDER_BASE_URL | http://localhost:11434/v1 | OpenAI-compatible endpoint | | PI_RAKIT_PROVIDER_API_KEY | unset | API key (resolved at request time) | | PI_RAKIT_PROVIDER_MODEL | llama3.2 | Model identifier | | PI_RAKIT_PROVIDER_MODEL_NAME | model identifier | Model display name | | PI_RAKIT_PROVIDER_CONTEXT_WINDOW | 128000 | Token context limit | | PI_RAKIT_PROVIDER_MAX_TOKENS | 8192 | Max output tokens | | PI_RAKIT_PROVIDER_REASONING | false | true/false or 1/0 | | PI_RAKIT_PROVIDER_IMAGES | false | Enable image input support |

export PI_RAKIT_PROVIDER_BASE_URL="https://api.example.com/v1"
export PI_RAKIT_PROVIDER_API_KEY="your-secret-key"
export PI_RAKIT_PROVIDER_MODEL="your-model-id"
pi

Do not commit API keys. The extension passes $PI_RAKIT_PROVIDER_API_KEY to Pi so Pi resolves the variable only when needed.

For the default local setup, ensure Ollama is running:

ollama pull llama3.2
ollama serve

Troubleshooting

• Model missing: restart Pi, run /rakit, and check the provider's model list.
• Discovery unavailable: verify GET <baseUrl>/models is supported, or use Enter model manually.
• Connection refused: start the local server and confirm the base URL includes /v1.
• Unauthorized: export PI_RAKIT_PROVIDER_API_KEY in the shell that starts Pi.
• Validation error: context window and max tokens must be positive integers; input type must be selected from the list.

Local development

pi -e ./packages/custom-provider/extensions/index.js