vision-mcp
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., "@vision-mcpDescribe what's in this image: /tmp/photo.png"
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.
vision-mcp
极简视觉 MCP server:让纯文本模型通过云端视觉模型"看图"。
所有工具都只返回文字——图片在云端视觉模型处理,模型侧永远只收到文本。 因此对不支持图像输入的模型(如 deepseek-v4-flash)100% 可用,也兼容任何支持视觉的模型。
工具
工具 | 作用 |
| 看图问答 / 描述(核心);支持视频自动抽帧 |
| 图片文字识别(原样返回) |
| 目标定位:返回 |
| ffprobe 探测视频/音频元数据(时长/分辨率/fps/编码/音轨) |
| 清空当前会话的视觉问答记忆 |
Related MCP server: glm-vision
内置机制
指数退避重试 — 429/5xx/网络错误自动等待后重试(免费 GLM 限流常态,稳定性关键); 主选模型(4.6 Flash)只重试 1~2 次,仍失败立即切兜底模型,避免反复撞限流拖长耗时
GLM 免费家族自动兜底 —
glm-4.6v-flash限流/超时 →glm-4v-flash(全程零费用)视频抽帧 — ffmpeg 动态 fps 抽帧(默认 ≤12 帧),逐帧分析后汇总;临时帧用完即清理
图片 base64 内容哈希缓存 —
~/.cache/vision-mcp/,同一文件不重复编码轻量会话记忆 — 进程内保留最近 N 轮问答(默认 4),多轮追问同一张图自动带上前文;
memory=false关闭、clear_memory()清空、VISION_MCP_MEMORY=0全局关
后端(provider)
provider | 模型 | 端点 | Key |
|
|
|
|
|
|
|
|
| 固定走 zhipu GLM 免费家族兜底链,不切 qwen | — | — |
自定义 | 传 | — | — |
key 读取顺序:环境变量 > ~/.qwen-mm-plugins/config > 报错。
运行
# 作为 MCP server(stdio),由 Agent 拉起:
python vision_server.py
# 依赖:仅 mcp + openai
pip install mcp openai任意支持 MCP 的 Agent 都能接入,只需注册一个 stdio server:
python /path/to/vision_server.py
ZCode 接入(已配置好)
单一插件市场 vision-mcp,插件名 vision-mcp,重启 ZCode 后即可用。
旧的双插件(qwen-mm-plugins-core / api)已移除,备份在
~/.zcode/cli/plugins/backup-qwen-mm-plugins-*。
使用示例(发给 agent 的话术)
看下这张截图:
vision_chat(images=["/path/to/s.png"], text="这是什么界面?列出全部文字")识别这张图片文字:
ocr(images=["/path/to/doc.png"])看这个视频:
vision_chat(videos=["/path/to/clip.mp4"], text="视频里发生了什么?")定位图中物体:
grounding(images=["/path/to/photo.png"], text="找到所有车辆")用千问看这张图:
vision_chat(images=[...], provider="qwen")
环境变量
变量 | 默认 | 说明 |
|
|
|
|
| 记忆轮数 |
|
| 视频抽帧上限(最大 64) |
注意
GLM 兜底链:4.6 Flash(glm-4.6v-flash) → V4 Flash(glm-4v-flash),全是免费模型。 显式传
glm-*模型也会挂上这条兜底链;qwen/自定义端点只试指定模型,不加别的。 整个流程默认只用免费 GLM 模型,不会调用 qwen 或其他模型。glm-4v-flash的max_tokens上限为 1024,server 会在兜底到它时自动钳制。图片支持本地路径(自动 base64)或 http(s) URL。
dry_run=true可预览请求结构(不真实请求,不消耗额度)。
This server cannot be deployed
Maintenance
Related MCP Connectors
Turn any LLM multimodal; generate images, voices, videos, 3D models, music, and more.
Image & PDF tools for AI agents: compress, convert, resize, PDF, AI vision, pipeline.
Generate AI images, video, voiceovers and music from Claude, ChatGPT or Cursor through 50+ models (Veo 3.1, Kling 3, Seedance, Nano Banana, GPT Image, ElevenLabs). Also image editing, upscaling, background removal, face swap, transcription, voice cloning and UGC-style video ads. Sign in with OAuth — no API key to paste. Tools are annotated (read-only vs. credit-spending); failed generations are refunded.
Image and video AI tools and your own pipelines, run from any AI assistant.
Related MCP Servers
- AlicenseAqualityDmaintenanceEnables AI agents to analyze images, extract text, compare images, and analyze video through any OpenAI-compatible vision model.484 npm20MIT
- FlicenseNot gradedqualityCmaintenanceMCP server that gives text-only models vision capabilities via free GLM vision models, supporting image description, OCR, chart/document analysis, and grounding with automatic model fallback.1-
- AlicenseNot gradedqualityCmaintenanceAdds image recognition and UI grounding capabilities to text-only LLMs through MCP tools, supporting local and cloud vision backends.24 npmMIT
- AlicenseNot gradedqualityCmaintenanceProvides vision capabilities to text-only LLMs via MCP, enabling image understanding, Q&A, OCR, and image processing through cloud multimodal APIs.MIT