Pi Provider - MiniMax

Note: This is an unofficial, community-built provider. It is not affiliated with, endorsed by, or connected to MiniMax.

A Pi coding agent provider extension for MiniMax AI's M2 and M3 series models, plus native Pi tools for MiniMax web search, image understanding, voice listing, and text-to-speech.

Features

• Anthropic API Compatible: Uses MiniMax's Anthropic-compatible endpoint for seamless integration
• Fast Model Updates: Intentionally replaces Pi's built-in minimax and minimax-cn catalogs so new MiniMax models can ship without waiting for a Pi release
• Extended Context: Up to 1,000,000 token context window on M3, 204,800 on M2 series
• Thinking Support: Built-in reasoning/thinking capabilities (adaptive for M3, budget-based for M2 series)
• Cost Tracking: Accurate token pricing for all models
• Native MiniMax Tools: Pi-native tools for MiniMax web search, image understanding, voice listing, and text-to-speech. No MCP server setup required.

Supported Models

Model	Context	Output	Input	Description
MiniMax-M3	1,000,000	up to 524,288	text, image	Frontier multimodal coding model with 1M context and adaptive thinking
MiniMax-M2.7	204,800	up to 65,536	text	Latest flagship with recursive self-improvement (~60 tps)
MiniMax-M2.7-highspeed	204,800	up to 65,536	text	Same as M2.7 with higher output speed (~100 tps)
MiniMax-M2.5	204,800	up to 65,536	text	Peak performance, ultimate value (~60 tps)
MiniMax-M2.5-highspeed	204,800	up to 65,536	text	Same as M2.5 with higher output speed (~100 tps)
MiniMax-M2.1	204,800	up to 65,536	text	Powerful multi-language programming (~60 tps)
MiniMax-M2.1-highspeed	204,800	up to 65,536	text	Same as M2.1 with higher output speed (~100 tps)
MiniMax-M2	204,800	up to 65,536	text	Agentic capabilities, advanced reasoning (~60 tps)

Notes:

• M3 supports image and video input directly. Only text and image are exposed to Pi today.
• M2 series do not accept image or video input; the provider only declares text.
• M2 series max_tokens field is configured to MiniMax's recommended cap (65,536). The hard API maximum is 204,800; the provider can pass that through if requested via Pi's per-request maxTokens option.

Installation

Option 1: Using pi install (npm)

pi install npm:@sinamtz/pi-minimax-provider

Option 2: Using pi install (git)

pi install git:https://github.com/sinamtz/pi-minimax-provider

Option 3: Install globally via npm

npm install -g @sinamtz/pi-minimax-provider

Option 4: Clone locally

git clone https://github.com/sinamtz/pi-minimax-provider.git
cd pi-minimax-provider

Usage

Set your API key

For MiniMax Token Plan / coding-tool usage, get a Subscription Key from the MiniMax Token Plan page. For pay-as-you-go Open Platform usage, use an Open Platform API key from Interface Key.

This provider sends the key as an Anthropic-compatible bearer token to https://api.minimax.io/anthropic by default, matching MiniMax's coding-tool configuration reference.

There are three ways to set your key (in order of priority):

Option 1: Using /login (Recommended)

The easiest way is to use pi's built-in login command:

pi -e ./pi-minimax-provider
/login

This will prompt you to enter your MiniMax key, which is then securely stored in ~/.pi/agent/auth.json.

Option 2: Edit auth.json manually

Add your MiniMax key directly to the auth file:

{
  "providers": {
    "minimax": {
      "access": "your-api-key-here",
      "expires": 1935753600000
    }
  }
}

Option 3: Environment variable

export MINIMAX_API_KEY=your-api-key-here
pi -e ./pi-minimax-provider

Select a region

The extension replaces both Pi MiniMax provider catalogs:

Provider	Base URL	Environment key
minimax	https://api.minimax.io/anthropic	MINIMAX_API_KEY
minimax-cn	https://api.minimaxi.com/anthropic	MINIMAX_CN_API_KEY

Select the corresponding provider/model in Pi. Optional MINIMAX_API_HOST and MINIMAX_CN_API_HOST variables can override either endpoint for proxies or custom routing.

Run pi with the extension

# Using local path
pi -e ./pi-provider-minimax

# Using a global install
pi -e $(npm root -g)/pi-provider-minimax

pi install npm:@sinamtz/pi-minimax-provider is the recommended way to install; Pi then loads the extension automatically.

Select a model

In pi's interactive mode:

/model MiniMax-M3

Or use the model selector UI. The provider registers the model id exactly as shown in Supported Models.

Native MiniMax Tools

Pi does not consume MCP servers directly, so this extension exposes MiniMax Token Plan capabilities as native Pi tools backed by MiniMax's official mmx-cli/sdk. The SDK is installed as a package dependency: users do not need a global mmx installation, MCP server, Python, or uvx.

Tool	Purpose
minimax_web_search	Performs MiniMax web search and returns organic results plus related searches.
minimax_understand_image	Analyzes JPEG, PNG, or WebP images from HTTP/HTTPS URLs, data URLs, or local file paths.
minimax_list_voices	Lists available MiniMax Token Plan speech voices.
minimax_text_to_audio	Generates speech audio using MiniMax voices, with URL or safe local-file output.
minimax_generate_image	Generates images and returns their URLs.
minimax_video	Generates videos, checks task status, and downloads completed videos.
minimax_generate_music	Generates music with optional lyrics or instrumental mode.
minimax_quota	Shows Token Plan quota and usage.

Tools resolve credentials through Pi's model registry first, so /login and auth.json work. They then fall back to MINIMAX_API_KEY or MINIMAX_CN_API_KEY. Set MINIMAX_REGION=global|cn to select a region explicitly; host overrides remain available through MINIMAX_API_HOST, MINIMAX_CN_API_HOST, or the official SDK's MINIMAX_BASE_URL.

Tool examples

Ask Pi for current information:

Use MiniMax web search to find recent MiniMax speech API documentation.

Analyze an image:

Use MiniMax image understanding to describe ./screenshot.png. Focus on visible UI text.

List voices:

Use MiniMax to list available system voices.

Generate speech, images, video, or music:

Use MiniMax text to speech to create an mp3 saying: Hello from Pi and MiniMax.
Generate a MiniMax image of a cyberpunk city at night.
Start a MiniMax video generation task showing ocean waves at sunset.
Generate an instrumental MiniMax jazz track.

Output handling

minimax_text_to_audio defaults to local file output and refuses to overwrite existing files unless replacement is explicitly requested. Set output_mode: "url" when link output is preferred and MiniMax supports it.

Cost

These capabilities consume MiniMax Token Plan quota. Use minimax_quota to inspect remaining usage. Voice cloning is not exposed because of consent, privacy, and cost considerations.

Prompt Caching

MiniMax supports automatic prompt caching for repeated prefixes such as tool lists, system prompts, and conversation history. The provider does not need to add MCP or explicit cache setup for this: MiniMax applies passive caching on supported models when requests have enough repeated input. The extension tracks cache read/write usage returned by the Anthropic-compatible stream so Pi cost accounting can reflect cache-hit tokens.

Explicit Anthropic cache_control is not enabled by default because MiniMax documents it for M2.x models, while M3 uses passive caching. This avoids adding model-specific cache markers that could be ignored or rejected.

Limitations

• M3 input above 512k tokens is in limited quantity and may require contacting MiniMax sales for access. The provider passes this through unchanged.
• M2 series do not support image or video input; the provider declares only text in the model schema.
• M2 series cache writes use a separate per-million-token rate ($0.375). M3 passive cache writes are not listed as a separate charge.

Pricing

Pricing is approximate and in USD per million tokens for pay-as-you-go standard tier:

Model	Input	Output	Cache Read	Cache Write
M3 ≤512k input tokens	$0.30	$1.20	$0.06	Not listed / no passive-cache write charge
M3 >512k input tokens	$0.60	$2.40	$0.12	Not listed / no passive-cache write charge
M2.7	$0.30	$1.20	$0.06	$0.375
M2.7-highspeed	$0.60	$2.40	$0.06	$0.375
M2.5	$0.30	$1.20	$0.03	$0.375
M2.5-highspeed	$0.60	$2.40	$0.03	$0.375
M2.1	$0.30	$1.20	$0.03	$0.375
M2.1-highspeed	$0.60	$2.40	$0.03	$0.375
M2	$0.30	$1.20	$0.03	$0.375

MiniMax may run temporary discounts or priority-tier pricing that differs from these baseline values. For the most current pricing, see the MiniMax pay-as-you-go pricing page.

API Details

• International Base URL: https://api.minimax.io/anthropic
• Mainland China Base URL: https://api.minimaxi.com/anthropic
• Provider behavior: intentionally replaces Pi's built-in minimax and minimax-cn model lists with this package's faster-moving catalog
• Tool endpoints: /v1/coding_plan/search, /v1/coding_plan/vlm, /v1/get_voice, /v1/t2a_v2
• Protocol: Anthropic Messages API compatible
• Authentication: Bearer token
  • For Token Plan / coding-tool usage, use a MiniMax Subscription Key
  • For pay-as-you-go usage, use an Open Platform API key
  • Priority: options.apiKey > auth.json (via /login) > environment variable MINIMAX_API_KEY
• Streaming: Full streaming support; usage events include input/output, cache read, and cache write token counts

Development

The package uses TypeScript and Vitest.

npm install
npm run check   # tsc --noEmit
npm run lint    # eslint
npm test        # vitest (watch mode)
npm run build   # tsc -p tsconfig.build.json -> dist/

Tests run against the published MiniMax HTTP shapes (mocked), so no live API calls are required and no costs are incurred. Do not commit real API keys.

Troubleshooting

"Model not found" error

• Ensure you've set up authentication (see Set your API key)
• Verify the extension path is correct when using -e
• Confirm the model id matches one of the Supported Models (case-sensitive)

Authentication errors

• Check your MiniMax API key is valid
• Ensure you have credits/quota available in your MiniMax account
• If using /login, re-authenticate: /logout -> minimax then /login -> minimax

Wrong region

• If requests time out or return auth errors, set MINIMAX_API_HOST to the correct regional endpoint (e.g. https://api.minimaxi.com for Mainland China)

Changelog

v1.1.6

• Replace both Pi minimax and minimax-cn catalogs with the extension's faster-moving model list
• Use MiniMax's official mmx-cli/sdk for native Token Plan tools
• Add image generation, video operations, music generation, and quota tools
• Resolve tool credentials through Pi's model registry, including /login and auth.json
• Update M3 pricing and Token Plan documentation from current MiniMax docs

v1.1.1

• Documentation: clarify supported models, output caps, region selection, prompt caching behavior, and add a development/test section

v1.1.0

• Add MiniMax-M3 model support with 1M context, image input, adaptive thinking, and documented 512K max output cap
• Add native Pi tools for MiniMax web search, image understanding, voice listing, and text-to-speech
• Track prompt-cache read/write usage returned by MiniMax
• Refresh model pricing, output limits, and README documentation from current MiniMax docs

v1.0.8

• Fix: Strip oauth: prefix from API keys to prevent 401 authentication errors
• Add vitest unit tests for API key cleaning

v1.0.7

• Add ESLint with TypeScript support
• Add lint script to npm run commands

v1.0.6

• Add auth.json support via /login minimax command

v1.0.3

• Initial release

License

MIT

Contributing

Contributions welcome! Please open an issue or PR on GitHub.