dsh-vision-dashscope
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., "@dsh-vision-dashscope describe this local video: ~/Downloads/sample.mp4"
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.
dsh-vision-dashscope
通过 DashScope / 千问(Qwen) 识别并生成图片、视频、音频的 DSH 插件(MCP server)。专治 Deepseek-omnimodal 等插件「视频大了就不行」的问题:
长视频/大视频直接识别:走百炼临时 OSS 直传(
oss://临时 URL,单文件上限 1GB,按模型的max_file_size_mb校验),不压缩、不损画质;小视频(≤14MB)走 base64 省一次上传。mkv / avi 等浏览器不可播格式也能识别(模型服务端解码抽帧)。
音频:短音频 qwen3.5-omni 直接理解,长音频自动 fun-asr 转写。
工具
工具 | 说明 |
| 本地图片 / 图片 URL → 描述或按任务分析(qwen4.5-flash) |
| 本地视频 / 视频 URL → 描述或按任务分析(qwen4.5-flash;大文件 OSS 直传) |
| 短音频 qwen3.5-omni 理解;长音频 fun-asr 转写 |
| 文生图(qwen-image-3.0 等;约 0.18 元/张起) |
| 文生视频(wan2.7-t2v / happyhorse-1.1-t2v;约 0.6 元/秒起) |
| 图生视频(wan2.7-i2v 等,首帧图片自动上传临时 OSS) |
| 查看当前模型与限制配置 |
识别工具会调用 DashScope 付费 API(按 token 计费,量小价低)。
生成工具必须
confirm=true才会实际调用付费接口;false时只返回预计费用(如“预计 3.00 元(5 秒 × 0.6 元/秒)”)。生成结果自动下载到本地输出目录(默认
~/Downloads/dsh-vision-dashscope,可用DASH_VISION_OUTPUT_DIR修改),返回本地路径——配合 dsh-image-preview 可直接在对话里内联预览。文生图默认额外转一张 JPEG 预览(
preview字段,最长边 1600px / 质量 85,可配置),原始 PNG 保留在files中——DSH 图片内联上限 5MB,大 PNG(1792×2400 可达 7MB)直接内联会 404,用 JPEG 预览即可正常显示。
Related MCP server: jimeng_visual_generation
安装
需要 Python 3.12+、uv。复制 .env.example 为 .env 并填写 API Key:
git clone https://github.com/ywleeo/dsh-vision-dashscope.git
cd dsh-vision-dashscope
cp .env.example .env # 填入 DASH_VISION_API_KEY
uv syncAPI Key 也可用环境变量:DASH_VISION_API_KEY(缺省回退 OMNIMODAL_API_KEY / DASHSCOPE_API_KEY)。
DSH 接入
在 ~/.dsh/profiles/web/cordis.patch.yml(用户 patch 层)追加:
- insert:
- id: mcp-vision-dashscope
name: '@deepseek-ai/dsh-mcp-client'
config:
serverName: vision-dashscope
transport: stdio
command: /Users/leeo/.local/bin/uv
args:
- run
- --directory
- /path/to/dsh-vision-dashscope
- dsh-vision-dashscope
failOnStartupError: false重启 dsh web。连接后工具以 mcp__vision-dashscope__recognize_video 等名字出现。
大文件上传原理(长视频关键)
GET /api/v1/uploads?action=getPolicy&model=<model>获取上传凭证(含upload_host、签名、max_file_size_mb);multipart 上传文件到 OSS 临时存储,得到
oss://<key>临时 URL(有效期 48h,免费);chat 请求
video_url: {"url": "oss://..."}并在 header 加X-DashScope-OssResourceResolve: enable,模型服务端拉取解码抽帧。
参考:上传文件获取临时 URL(阿里云百炼文档)、视觉理解(视频)。
配置(环境变量 / .env)
变量 | 默认 | 说明 |
| — | API Key(回退 OMNIMODAL_API_KEY / DASHSCOPE_API_KEY) |
|
| 图片识别模型 |
|
| 视频识别模型 |
|
| 短音频理解模型 |
|
| 长音频转写模型 |
|
| 视频 base64 阈值,超出走 OSS 直传 |
|
| 长音频转写阈值(秒) |
|
| 文生图模型(PRO/MAX 档同理) |
|
| 文生视频模型(MAX 档 happyhorse-1.1-t2v;I2V 同理) |
|
| 生成结果保存目录 |
|
| 视频生成时长上限(秒) |
|
| 输出上限 |
测试
uv run pytest -qLicense
MIT
Available Tools
7 toolsdashscope_statusA
查看当前模型与限制配置(不含 API Key)。
Inputs: no named parameters. Returns: result
generate_imageA
文生图。生成结果下载到本地输出目录并返回路径(配合 dsh-image-preview 可内联预览)。
Inputs: prompt (required), size, tier, confirm.
generate_videoA
文生视频。生成结果下载到本地输出目录并返回路径。
Inputs: prompt (required), tier, wait, confirm, +2 more.
generate_video_from_imageB
图生视频:以图片为首帧生成视频。
Inputs: image (required), prompt (required), tier, wait, +3 more.
recognize_audioA
识别本地音频或音频 URL。
短音频(≤300s 且 ≤10MB)走 qwen3.5-omni 直接理解内容;长音频自动走 fun-asr 异步转写(本地文件先上传到临时 OSS),返回纯文本转录。
Inputs: audio (required), task, language. Returns: result
recognize_imageA
识别本地图片或图片 URL,按 task 调用千问多模态模型。
Inputs: image (required), mode, task. Returns: result
recognize_videoA
识别本地视频或视频 URL。
大文件(>14MB)自动走百炼临时 OSS 直传(上限 1GB、不压缩画质),小文件走 base64;支持任意浏览器不可播放的容器(mkv/avi 等),由模型服务端解码抽帧。
Inputs: video (required), fps, mode, task. Returns: result
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 7 tools
Each tool targets a distinct modality and action: recognize_image/video/audio are clearly separate input types, while generate_image/video/video_from_image are distinct generation tasks. dashscope_status is the only non-media utility, so there is no real ambiguity.
The recognize_* and generate_* groups follow a consistent verb_noun underscore pattern. dashscope_status is the only outlier since it uses a noun-style name rather than verb_noun, but the overall convention is still predictable and readable.
Seven tools is well-scoped for a vision/media DashScope server. Each tool covers a meaningful capability without redundant or excessive surface area.
The tool set covers both major media understanding paths (image, video, audio) and generation paths (image, video, image-to-video). The status tool fills the configuration/inspection need, and there are no obvious dead ends for the stated domain.
Maintenance
Related MCP Connectors
MCP server for Wan AI video generation
MCP server for Qwen Image 3 AI image generation
MCP server for Hailuo (MiniMax) AI video generation
MCP server for ByteDance Seedream AI image generation
Related MCP Servers
- FlicenseNot gradedqualityCmaintenanceMCP server enabling AI agents to generate AI videos and images, analyze video content, and download videos from Douyin and Xiaohongshu.-
- AlicenseAqualityCmaintenanceMCP server for generating images and videos using Volcengine's Jimeng APIs, supporting text-to-image, image-to-image, multi-image fusion, text-to-video, and image-to-video.31MIT
- AlicenseAqualityDmaintenanceMCP server for AI-powered media generation: images, videos, audio, and upscaling using 99 AI models.6MIT
- FlicenseNot gradedqualityDmaintenanceAn MCP server for generating images and videos via external multimodal models, providing unified tools like imageGenerate, videoGenerate, and materialWrite with support for normal and smart generation modes.-