Skip to main content
Glama
MiniMax MCP Server

MiniMax MCP Server

by longhz

MiniMax MCP Server

A fully functional MiniMax Token Plan MCP Server, providing four capabilities: quota query, web search, image understanding, and image generation. Built on Python FastMCP, it can be directly integrated into MCP-supported clients like Claude Code.

Feature Overview

Tool

Function

MiniMax API Used

query_quota

Query remaining quota for Token Plan models

/v1/token_plan/remains

web_search

MiniMax web search

/v1/coding_plan/search

understand_image

Image understanding (dual mode + cache)

/v1/coding_plan/vlm

generate_image

Image generation (image-01 model)

/v1/image_generation

clear_vision_cache

Clear vision analysis cache

-

Related MCP server: minimax-mcp

Project Structure

minimax-mcp/
├── config.env.example       # 配置模板
├── pyproject.toml
├── src/minimax_mcp/
│   ├── server.py            # FastMCP 服务入口 + 5 个 Tool
│   ├── client.py            # MiniMax HTTP API 客户端
│   ├── config.py            # 配置管理(config.env → 环境变量 → 默认值)
│   ├── tools/
│   │   ├── quota.py         # 额度查询(5小时/日周期分类)
│   │   ├── web_search.py    # 网页搜索
│   │   ├── image_understand.py  # 图片理解(本地文件→Base64 转换)
│   │   └── image_generate.py    # 图片生成(自动保存本地)
│   └── vision/
│       ├── analyzer.py      # 分析器(双模式 + 缓存调度)
│       ├── prompts.py       # 结构化 Prompt 模板
│       └── cache.py         # LRU 内存缓存 + 磁盘 JSON 持久化
└── tests/

Quick Start

1. Get API Key

Obtain your API Key from the MiniMax Platform Token Plan page.

2. Installation

# 克隆仓库
git clone <repo-url>
cd minimax-mcp

# 安装依赖(需要 Python >= 3.10)
pip install -e .

# 或使用 uv
uv pip install -e .

3. Configuration

cp config.env.example config.env
# 编辑 config.env,填入你的 API Key

config.env structure:

# 必需
MINIMAX_API_KEY=你的_api_key

# 可选
MINIMAX_API_HOST=https://api.minimaxi.com          # 中国大陆
# MINIMAX_API_HOST=https://api.minimax.io          # 全球
MINIMAX_IMAGE_OUTPUT_DIR=~/Pictures/MiniMax         # 生成图片保存位置
MINIMAX_CACHE_DIR=~/.minimax-mcp/cache              # 视觉分析缓存位置
MINIMAX_VISION_DEFAULT_MODE=detailed                # quick 或 detailed
MINIMAX_CACHE_MAX_SIZE=256                          # 最大缓存条目
MINIMAX_CACHE_TTL_DAYS=7                            # 缓存过期天数

Important: config.env contains the API Key and is excluded in .gitignore. Do not commit to Git.

4. Register to Claude Code

Add to your Claude Code MCP configuration:

{
  "mcpServers": {
    "MiniMaxMCP": {
      "command": "uv",
      "args": ["run", "--directory", "<项目路径>/minimax-mcp", "minimax-mcp"]
    }
  }
}

Or use the system Python directly:

{
  "mcpServers": {
    "MiniMaxMCP": {
      "command": "python",
      "args": ["-X", "utf8", "-m", "minimax_mcp.server"],
      "env": {
        "PYTHONPATH": "<项目路径>/minimax-mcp/src"
      }
    }
  }
}

Tool Details

1. query_quota — Quota Query

Query the model quota usage of the Token Plan, automatically distinguishing between text models (5-hour cycle) and other models (daily cycle).

query_quota()

→ {
    text_models: [
      { model_name: "MiniMax-M3.5", used: 32, total: 100, remaining: 68, usage_pct: 32 }
    ],
    other_models: [
      { model_name: "image-01", used: 5, total: 50, remaining: 45, usage_pct: 10 }
    ],
    no_quota: [...],
    summary: "文本模型: 5:00:00 后重置 | 其他模型: 12:30:00 后重置"
  }

2. web_search — Web Search

Search for web content via the MiniMax search engine.

web_search(query="OpenAI GPT-5 发布日期")

→ {
    success: true,
    results: [
      { title: "...", url: "...", snippet: "...", position: 1 },
      ...
    ],
    related_searches: [...],
    query: "OpenAI GPT-5 发布日期"
  }

3. understand_image — Image Understanding

Drawing on the design philosophy of OpenHanako Vision Bridge, it provides two analysis modes.

Design Philosophy:

  • Offload images to a specialized vision model (MiniMax VLM) for structured analysis

  • Inject analysis results as text into the LLM context, allowing pure text models to "understand" image content

  • Built-in LRU cache (256 entries, disk-persistent); identical image + identical prompt does not consume extra quota

Dual Modes:

Mode

Use Case

Output Format

quick

Quickly understand image content

~300-word concise description

detailed (default)

In-depth analysis

7-dimension structured report

detailed mode output dimensions:

Dimension

Content

image_overview

Image overview

visible_text

Visible text

objects_and_layout

Objects and layout

charts_or_data

Charts/data

answer_to_request

Answer to user question

evidence

Analysis evidence

uncertainty

Uncertainty notes

understand_image(
    image_url="https://example.com/photo.jpg",   # 支持 HTTP URL 或本地路径
    prompt="图片里有什么错误提示?",                # 可选,特定问题
    mode="detailed",                              # quick 或 detailed
    use_cache=true                                # 默认启用缓存
)

Note: MiniMax VLM does not support coordinate output, so OpenHanako's Visual Primitives spatial coordinate annotation capability is not yet implemented. If coordinate-aware image analysis is required, it is recommended to use a model that supports visual primitives.

4. generate_image — Image Generation

Generate images using the MiniMax image-01 model.

generate_image(
    prompt="A serene lake at sunset with snow-capped mountains",
    model="image-01",           # 目前仅支持 image-01
    aspect_ratio="16:9",        # 1:1 / 16:9 / 9:16 / 3:4 / 4:3
    n=1,                        # 1-3 张
    prompt_optimizer=true,      # 启用提示词自动优化
    save_to_disk=true,          # 自动保存到本地
    response_format="base64"    # base64(可存本地)或 url(24h临时链接)
)

Images are automatically saved to the MINIMAX_IMAGE_OUTPUT_DIR directory (default ~/Pictures/MiniMax).

Architecture Design

┌─────────────────┐     MCP Protocol     ┌──────────────────────┐
│  Claude Code /   │ ◄──────────────────► │  FastMCP Server       │
│  MCP Client      │     (stdio)          │  (server.py)          │
└─────────────────┘                       │                       │
                                          │  ┌─────────────────┐ │
                                          │  │ quota.py        │ │
                                          │  │ web_search.py   │ │
                                          │  │ image_*.py      │ │
                                          │  │ vision/analyzer │ │
                                          │  │ vision/cache    │ │
                                          │  └────────┬────────┘ │
                                          │           │           │
                                          │  ┌────────▼────────┐ │
                                          │  │ MiniMaxClient   │ │
                                          │  │ (HTTP/HTTPS)    │ │
                                          │  └────────┬────────┘ │
                                          └───────────┼───────────┘
                                                      │
                                          ┌───────────▼───────────┐
                                          │  MiniMax API          │
                                          │  api.minimaxi.com     │
                                          └───────────────────────┘

Configuration Priority

config.env 文件 → 环境变量 → 代码默认值

Caching Mechanism (Vision Analysis)

  • Memory Cache: LRU strategy, max 256 entries (configurable)

  • Disk Persistence: JSON format, stored in MINIMAX_CACHE_DIR

  • Cache Key: SHA256(image_url + prompt + mode)

  • TTL: 7-day expiration by default

Dependencies

  • Python >= 3.10

  • mcp >= 1.0.0 (FastMCP / MCP Protocol)

  • httpx >= 0.27.0 (HTTP client)

  • Pillow >= 10.0.0 (Image processing)

Development

# 克隆并安装开发依赖
git clone <repo-url>
cd minimax-mcp
pip install -e ".[dev]"

# 运行测试
python -m pytest tests/

# 直接启动 MCP 服务器
python -m minimax_mcp.server

References

License

MIT

Available Tools

5 tools
clear_vision_cacheA

清除图片理解缓存(手动触发)

将内存中的缓存数据强制写入磁盘,并返回当前缓存统计。

Inputs: no named parameters.

generate_imageA

使用 MiniMax image-01 生成图片,默认保存到本地

图片自动保存到配置的输出目录(默认 ~/Pictures/MiniMax)。

Args: prompt: 图片描述文本(英文效果最佳,中文也可) model: 模型名称,默认 image-01 aspect_ratio: 图片比例:1:1 / 16:9 / 9:16 / 3:4 / 4:3,默认 1:1 response_format: 默认 base64(自动存本地),改 url 则仅返回24h临时链接 n: 生成数量 1-3,默认 1 prompt_optimizer: 是否启用提示词优化,默认 true save_to_disk: 是否保存到本地(仅 base64 模式有效),默认 true

Inputs: prompt (required), model, aspect_ratio, response_format, +3 more.

query_quotaA

查询 MiniMax Token Plan 剩余额度

返回文本模型(5小时周期)和其他模型(日周期)的配额使用情况, 包括已用次数、剩余次数、使用百分比、重置时间等。

无需参数,使用 config.env 中配置的 API Key。

Inputs: no named parameters.

understand_imageA

分析图片内容 — 借鉴 OpenHanako Vision Bridge 设计

两种模式:

  • quick: 简洁描述(~300词),适合快速了解图片内容

  • detailed: 结构化 7 维度分析(image_overview / visible_text / objects_and_layout / charts_or_data / answer_to_request / evidence / uncertainty),参考 OpenHanako vision-bridge.js

支持缓存:同一图片+相同prompt不重复调用API(LRU + 磁盘持久化)

Args: image_url: 图片 URL(HTTP/HTTPS)或本地文件路径,支持 JPEG/PNG/GIF/WebP (≤20MB) prompt: 对图片的具体问题,如 "这张图片里有什么错误提示?" mode: "quick" 或 "detailed",默认 detailed use_cache: 是否使用缓存,默认 true

Inputs: image_url (required), prompt, mode, use_cache.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A4.2/5.0

    Scored across 5 tools

    Disambiguation5/5

    Each tool targets a distinct operation (cache clearing, image generation, quota query, image analysis, web search) with no overlap in purpose.

    Naming Consistency5/5

    All tool names follow a consistent verb_noun snake_case pattern (e.g., clear_vision_cache, generate_image, query_quota, understand_image, web_search).

    Tool Count5/5

    Five tools is a well-scoped set for a server offering image generation, analysis, web search, cache management, and quota checking—each tool serves a clear and necessary function.

    Completeness5/5

    The tool set covers the core functionalities of the MiniMax AI service: image generation, image understanding, web search, cache management, and quota monitoring, with no obvious gaps.

    Related MCP Connectors

    Related MCP Servers

    • F
      license
      Not graded
      quality
      C
      maintenance
      Exposes Qwen AI Token Plan's harness tools (web search, web extraction, code interpreter, text-to-image, and image-to-image search) as MCP tools. It enables agents to perform real-time web searches, extract webpage content, execute code in a sandbox, and search images by text or image queries.
      -
    • F
      license
      Not graded
      quality
      A
      maintenance
      Enables MCP clients such as Claude Desktop, Cursor, and Windsurf to perform web searches and analyze images through MiniMax APIs, returning structured search results and AI-generated answers about images from URLs or local files. It is tailored for coding-plan users working in code development workflows.
      -