# MiniMax Vision MCP Server

MCP server for **MiniMax** vision models — analyze images through the
OpenAI-compatible `POST /v1/chat/completions` endpoint.

## Features

- 🖼️ Analyze images — local files (png/jpg/jpeg/gif/webp/bmp) **and remote HTTP(S) URLs**
- 🧠 **MiniMax-M3**: image + video understanding, 1M context, adaptive thinking
- 🎛️ `temperature` fully configurable `[0, 2]` (unlike some providers that lock it)
- 💭 `thinking` flag → enables adaptive thinking + `reasoning_split` on M3
- ⚡ Zero non-MCP dependencies, one-line `npx` deploy

## Requirements

- Node.js >= 18
- A [MiniMax platform](https://platform.minimaxi.com) API key (账户管理 → 接口密钥)

## Install & run

```bash
cd minimax-vision-mcp-server
npm install
npm start
```

## Environment variables

| Variable | Required | Default | Description |
|---|---|---|---|
| `MINIMAX_API_KEY` | ✅ | — | Your MiniMax API key. |
| `MINIMAX_MODEL` | | `MiniMax-M3` | Model: `MiniMax-M3`, `MiniMax-M2.7`, `MiniMax-M2.7-highspeed`, `MiniMax-M2.5`, `MiniMax-M2.5-highspeed`, `MiniMax-M2.1`, `MiniMax-M2.1-highspeed`, `MiniMax-M2`. |
| `MINIMAX_BASE_URL` | | `https://api.minimax.chat/v1` | Override endpoint (for proxies). |
| `MINIMAX_MAX_TOKENS` | | `8192` | Default max output tokens. M3 supports up to 524288. |

> Note: International users may use `https://api.minimaxi.com/v1` as the base URL.

## Claude Code / CC-Switch config

```json
{
  "mcpServers": {
    "minimax-vision": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "minimax-vision-mcp-server"],
      "env": {
        "MINIMAX_API_KEY": "your-minimax-api-key",
        "MINIMAX_MODEL": "MiniMax-M3"
      }
    }
  }
}
```

## Tool: `minimax_vision_understand`

| Parameter | Type | Required | Description |
|---|---|---|---|
| `image` | string | ✅ | Local image path **or** remote HTTP(S) URL. |
| `prompt` | string | ✅ | What to ask about the image. |
| `max_tokens` | number | | Max output tokens. Default 8192. |
| `temperature` | number | | 0-2, default 1. |
| `thinking` | bool | | Enable adaptive thinking + reasoning split (M3). M2.x always think; ignored there. |

## Why MiniMax for vision?

- **1M context** on M3 — analyze long documents alongside images
- **Native image + video understanding**
- **Configurable temperature** — fine-grained control over determinism
- **Adaptive thinking** — reasoning on by default, splittable via `reasoning_split`

## Related projects

- [kimi-vision-mcp-server](https://github.com/kira4094/kimi-vision-mcp-server) — Moonshot Kimi
- [doubao-vision-mcp-server](https://github.com/kira4094/doubao-vision-mcp-server) — ByteDance Doubao
- [glm-vision-mcp-server](https://github.com/kira4094/glm-vision-mcp-server) — Zhipu GLM
- [qwen-vision-mcp-server](https://github.com/kira4094/qwen-vision-mcp-server) — Alibaba Qwen
- [@kira4094/agnes-image-mcp-server](https://www.npmjs.com/package/@kira4094/agnes-image-mcp-server) — Agnes Image (text-to-image)

## License

MIT