🪐 pi-tarmis-provider

Open models on Tarmis's own GPUs for pi

One OpenAI-compatible API to open models on Tarmis's own GPUs — DeepSeek, GLM, Kimi, MiniMax, and more.

Features

• OpenAI-compatible API — Uses Tarmis's /v1/chat/completions endpoint
• Rich auto-synced catalogue — The Tarmis /v1/models endpoint returns per-model pricing, context windows, max output, modalities, and supported features; models.json is generated directly from it
• Public model list — /v1/models is unauthenticated, so pi lists Tarmis models (and the background model sync runs) even before you set a key
• Reasoning models — Thinking models are detected from the API's supported_features and configured for reasoning_effort control
• Vision models — Image input is auto-detected from input_modalities
• Tool use — Function calling for models that advertise tools support
• Streaming — Real-time token streaming
• Prompt caching — Cached input is billed at a cheaper rate when the API reports an input_cache_read price

Available Models

| Model | Context | Vision | Reasoning | Input $/M | Cache Read $/M | Output $/M | |-------|---------|--------|-----------|-----------|-----------------|------------| | DeepSeek V4 Flash | 1.0M | ❌ | ✅ | $0.09 | $0.02 | $0.18 | | DeepSeek V4 Pro | 1.0M | ❌ | ✅ | $0.43 | $0.09 | $0.87 | | GLM-5.2 | 1.0M | ❌ | ✅ | $0.60 | $0.14 | $2.20 | | Kimi K3 | 1.0M | ✅ | ✅ | $2.50 | $0.25 | $10.00 | | MiniMax M3 | 524K | ❌ | ✅ | $0.30 | $0.06 | $1.20 |

The table above is regenerated by node scripts/update-models.js from the live Tarmis API. Costs are per million tokens (the API reports per-token prices). Prices subject to change — check docs.tarmis.ai/docs/pricing for current pricing.

Installation

Option 1: Using pi install (Recommended)

Install from npm (published as pi-tarmis-provider) or directly from GitHub:

# From npm
pi install npm:pi-tarmis-provider

# Or directly from GitHub
pi install https://github.com/monotykamary/pi-tarmis-provider

Then set your API key and run pi:

# Recommended: add to auth.json
# See Authentication section below

# Or set as environment variable
export TARMIS_API_KEY=your-api-key-here

pi

Get your API key from tarmis.ai.

Option 2: Manual Clone

1. Clone this repository:

   git clone https://github.com/monotykamary/pi-tarmis-provider.git
   cd pi-tarmis-provider

2. Set your Tarmis API key:

   # Recommended: add to auth.json
   # See Authentication section below
   
   # Or set as environment variable
   export TARMIS_API_KEY=your-api-key-here

3. Run pi with the extension:

   pi -e /path/to/pi-tarmis-provider

Authentication

The Tarmis API key can be configured in multiple ways (resolved in this order):

1. auth.json (recommended) — Add to ~/.pi/agent/auth.json:

   { "tarmis": { "type": "api_key", "key": "your-api-key" } }

   The key field supports literal values, env var names, and shell commands (prefix with !). See pi's auth file docs for details.
2. Runtime override — Use the --api-key CLI flag
3. Environment variable — Set TARMIS_API_KEY

Get your API key from tarmis.ai.

Note: A key is only required to actually call a model. The model catalogue (/v1/models) is public, so pi displays available Tarmis models and keeps them synced even before a key is configured.

Environment Variables

| Variable | Required | Description | |----------|----------|-------------| | TARMIS_API_KEY | No | Your Tarmis API key (fallback if not in auth.json; only needed to call a model) |

Configuration

Add to your pi configuration for automatic loading:

{
  "extensions": [
    "/path/to/pi-tarmis-provider"
  ]
}

Usage

Once loaded, select a model with:

/model tarmis minimax/minimax-m3

Or use /models to browse all available Tarmis models.

Reasoning Effort

For reasoning models, control thinking depth:

/reasoning high

Values: none, low, medium, high, max

API Compatibility

Tarmis provides an OpenAI-compatible API. Key notes:

| Aspect | Behavior | |--------|----------| | Endpoint | https://api.tarmis.ai/v1/chat/completions | | Max tokens field | Both max_tokens and max_completion_tokens accepted | | Thinking format | openai (reasoning_content in response, reasoning_effort in request) | | Developer role | Not assumed (developer messages are converted to system) | | Reasoning effort | Accepted by reasoning models | | /v1/models | Public (no auth) — returns pricing, context, modalities, and features | | Prompt caching | Cached input billed at a cheaper input_cache_read rate when the API exposes one |

Reasoning compat: The Tarmis /v1/models endpoint reports which models support reasoning, but not the wire format each model expects. Reasoning models are given the standard OpenAI-compatible defaults (thinkingFormat: "openai", supportsReasoningEffort: true, supportsDeveloperRole: false). If a model on Tarmis needs a different format (for example DeepSeek's native thinking field), override it in patch.json — see AGENTS.md.

API Documentation

• Tarmis Docs: https://docs.tarmis.ai
• OpenAI-compatible endpoint: https://api.tarmis.ai/v1
• Models endpoint (public): https://api.tarmis.ai/v1/models
• Pricing: https://docs.tarmis.ai/docs/pricing

Updating Models

Run the update script to fetch the latest models from Tarmis's API:

node scripts/update-models.js

The /v1/models endpoint is public, so no API key is required. This will:

1. Fetch models from https://api.tarmis.ai/v1/models
2. Derive pricing, context windows, max output, vision, and reasoning support from the API metadata
3. Apply overrides from patch.json
4. Update models.json and the README model table
5. Move delisted models into deprecated-models.json (14-day grace period)

License

MIT