mcp-multimedia-server
Provides tools for analyzing images, videos, audio, and performing speech-to-text using Xiaomi's MiMo V2.5 multimodal model, with local preprocessing for compression and encoding.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@mcp-multimedia-server Analyze this video and summarize its content."
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
MCP Multimedia Server
给 LLM agent(Claude Code / Codex 等)外挂小米 MiMo V2.5 的多模态理解能力——图片、视频、纯音频、语音转写(ASR),全部走云端 MiMo V2.5 完成,本地只做必要的预处理(图片压缩、视频重编码)以省 token / 控制体积。
MCP(MCP = Model Context Protocol)stdio 服务器,注册进 agent 后即可当工具直接调用。
用途
图片理解 / OCR:单张或批量(一次最多 100 张,适合文档逐页解析)
视频理解:本地视频自动压缩重编码后发送,
fps可调精细度音频理解 / STT:转录或总结一段语音
语音转写(ASR):纯转写,支持语种指定
底层统一使用 MiMo V2.5(原生多模态,1M 上下文),通过任意 OpenAI 兼容端点接入
Related MCP server: Vision MCP Server
工具
工具 | 输入 | 说明 |
|
| 描述/OCR/图表/UI/报错诊断;本地图片默认压缩(8MP 上限 + A4-OCR 分辨率, |
|
| 本地视频自动重编码(短边 720p);支持 URL |
|
| 音频理解/STT(转录+总结);推荐本地文件或 base64 |
|
| 纯语音转写;仅 wav/mp3,≤10MB |
| 本地路径 | 读取图片尺寸/格式/EXIF/GPS,不调用云端 |
| — | 查看当前模型与各项限制/参数 |
工作原理
agent ── MCP stdio ──> 本服务 ── OpenAI 兼容端点 ──> MiMo V2.5(云端)
│
├─ 图片:本地压缩(8MP 上限 + A4-OCR 分辨率)→ image_url
├─ 视频:ffmpeg 重编码(短边720p + fps)→ video_url
└─ 音频:本地/base64 → input_audio(data + format)本地处理仅在发送前预处理,不做任何媒体理解
所有推理/识别/转写在云端 MiMo V2.5 完成
上下文窗口 1M tokens;媒体按分辨率/时长计 token(压缩后 A4 图约 1.8K token/张)
环境要求
Python 3.10+
ffmpeg(视频重编码用,须在 PATH)一个 OpenAI 兼容的多模态端点 + API key(如小米 MiMo 官方
api.xiaomimimo.com,或任意代理网关)
安装
# 1. 建 venv 并从 GitHub 安装
uv venv ~/.local/venvs/mcp-multimedia-server
uv pip install --python ~/.local/venvs/mcp-multimedia-server \
git+https://github.com/ThermalEng/mcp-multimedia-server.git
# 2. ffmpeg(视频重编码)
sudo apt install ffmpeg配置
操作参数(大小上限、压缩目标、视频重编码、格式白名单等)硬编码在 mcp_multimedia_server/config.py,改代码即可调整。上游三件套从进程环境读取:
环境变量 | 说明 |
| OpenAI 兼容端点(默认 |
| API key(不写进代码/仓库,只放环境) |
| 模型名(默认 |
Claude Code 注册(stdio)
把下面合并到 ~/.claude.json 的 mcpServers:
{
"mcpServers": {
"multimedia": {
"type": "stdio",
"command": "~/.local/venvs/mcp-multimedia-server/bin/mcp-multimedia-server",
"args": [],
"env": {
"MCP_MEDIA_BASE_URL": "https://opencode.ai/zen/go/v1",
"MCP_MEDIA_API_KEY": "sk-你的key",
"MCP_MEDIA_MODEL": "mimo-v2.5"
}
}
}
}注册后重启 agent,即可用 analyze_image / analyze_video / analyze_audio / asr 等工具。
备注
依赖
mcpSDK 须钉在<2(SDK 2.x 移除了Server.list_tools装饰器)视频重编码依赖系统
ffmpeg(config.VIDEO_FFMPEG可指绝对路径)专用模型说明:
mimo-v2.5-asr/mimo-v2.5-tts为官方独立模型;若你的端点未开放它们(部分网关 401),ASR 会用mimo-v2.5顶替(仍可转写),TTS 则不可用流式/SSE 传输:
MCP_TRANSPORT=sse/MCP_PORT=8093(需pip install mcp-multimedia-server[sse])
Available Tools
6 toolsanalyze_audioA
音频理解/STT:转录或总结一段音频内容。推荐用本地文件或 base64(URL 音频可能不被识别)。不传 prompt 默认转录并总结主旨。
Inputs: audio (required), prompt, max_tokens, temperature.
analyze_imageA
图片理解/OCR:分析一张或多张图片,支持描述、文字识别(OCR)、图表/UI/报错诊断等。单张用 image,多张(如文档多页)用 image_batch 数组。可用 preset 选任务类型(describe/ocr/chart/ui/diagram/error)或自由写 prompt。本地图片自动压缩优化(OCR 足够清晰、更快更省),需原图细节可设 compress=false。
Inputs: image, preset, prompt, compress, +3 more.
analyze_videoA
视频理解:让模型描述视频内容、按时间顺序总结。支持本地文件、URL 或 base64,本地视频自动压缩。fps 控制抽帧密度(默认 2,范围 1-30;越大时序越精细、越耗 token),一般用默认即可。
Inputs: video (required), fps, prompt, max_tokens, +2 more.
asrA
语音转写(ASR):把音频转成纯文本,适合会议记录、方言、嘈杂环境录音。仅支持 wav/mp3,最大 10MB。用 language 明确语种(auto/zh/en)可提高准确率。
Inputs: audio (required), language, max_tokens.
get_server_statusA
查询服务状态:云端模型是否已配置、当前模型与各项限制/处理参数。
Inputs: no named parameters.
image_metadataC
读取本地图片的元信息(尺寸、格式、颜色模式、EXIF/GPS)。需本地文件路径,不调用云端。
Inputs: image (required).
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 6 tools
Most tools are clearly distinct by media type (video, audio, image, status). The main overlap is between analyze_audio and asr, both handling audio transcription, but asr is specifically pure text with format/size limits, while analyze_audio offers broader understanding and summarization, making the distinction functional.
There is a strong pattern with analyze_video, analyze_audio, analyze_image, but asr and image_metadata break the verb_noun convention. Mixed styles (acronym, noun_phrase) and vague naming like asr make it less predictable.
Six tools cover the essential multimedia operations (audio/video/image analysis, speech transcription, metadata, status) without redundancy. The count is well-scoped for a multimedia server, each tool serving a clear purpose.
The tool surface covers major analysis needs: video, audio, image understanding, pure transcription, and metadata for images. Minor gaps exist like no metadata extraction for audio/video or a dedicated video transcription tool, but these are not critical for the apparent domain.
Maintenance
Related MCP Connectors
Turn any LLM multimodal; generate images, voices, videos, 3D models, music, and more.
Video, audio, and image processing for AI agents: convert, transcribe, upscale - 150+ operations.
OCR, transcription, file extraction, and image generation for AI agents via MCP.
214 tools for AI agents over one MCP endpoint: image, video, audio, 3D, data, storage, mail, GPUs.
Related MCP Servers
- AlicenseNot gradedqualityDmaintenanceProvides an OpenAI-compatible gateway to the MiMo (Xiaomi) model series, with native MCP server tools for web search and visual analysis.30GPL 3.0
- AlicenseAqualityDmaintenanceEnables AI agents to analyze images, extract text, compare images, and analyze video through any OpenAI-compatible vision model.4330 npm20MIT

cerase-media MCPofficial
AlicenseNot gradedqualityBmaintenanceProvides multimodal understanding tools including OCR, image description, audio transcription, UI screenshot analysis, and screenshot comparison via async tools on a multimodal endpoint.MIT- AlicenseAqualityCmaintenanceIntegrates Xiaomi MiMo's multimodal API to enable understanding of images, audio, and video through natural language prompts.3MIT