Keep a text-only DeepSeek model for conversation while a VLM transcribes images. Remote endpoints receive image bytes; local Ollama is an alternative.
DSH integration
Native runtime
Author-claimed
Safety audit
Unaudited
Last verified
2026-10-18
License
MIT
01What can it help you accomplish?
Use attached images while retaining a text-only DeepSeek conversation model
VLM-transcribed image text supplied to DeepSeek for its answer
DSH GUI users choosing a text-only model
02How to install into DeepSeek Harness
Prerequisites
- DeepSeek Harness; the README reports adaptation to dsh 0.1.1-rc.2. Node >=22.19.
- Provide a VLM API key or a local Ollama service with a vision model. Keys can come from apiKey, VISION_API_KEY or DASHSCOPE_API_KEY.
Installation steps
- 01
$ dsh plugin --profile web add dsh-vision-proxy
- 02
If pnpm reports ignored build scripts, add dsh-vision-proxy: true and sharp: true under allowBuilds in the profile pnpm-workspace.yaml, then rerun the install command.
- 03
Answer the installer prompt: y for a VLM API key, or N for the local path.
- 04
Restart dsh web and choose DeepSeek + 自动识图 in the model selector.
Verify the integration
- Paste an image into a conversation; look for [图片转译] followed by the DeepSeek answer.
03DSH integration and capability boundaries
Native deepseek-vision provider route transcribing attached images before calling DeepSeek
Attached-image transcription
GUI image attachments→Image descriptions converted to text for DeepSeek
Image bytes go to the configured VLM endpoint, unless it is a local service.In-process result cache
Image content hashes→Reuse of image transcription results
Transcriptions are cached in process memory and not persisted to disk.
04Who is it for? When not to use it?
Good for
- DSH GUI users choosing a text-only model
Not for
- Image bytes leave the machine when a remote VLM endpoint is configured. Use a local endpoint such as Ollama to keep that transcription request local.
- Dense UI screenshots may lose small text details; the visual model limits transcription quality.
05Compatibility, maintenance and safety notes
- Image bytes leave the machine when a remote VLM endpoint is configured. Use a local endpoint such as Ollama to keep that transcription request local.
- Dense UI screenshots may lose small text details; the visual model limits transcription quality.
README claims; MIT license
06Frequently asked questions
Which DSH model route should I choose?
After restarting dsh web, choose DeepSeek + 自动识图 in the model selector.
Can DSH image transcription stay on my machine?
Yes, when the VLM baseURL points to a local service such as Ollama. A remote endpoint receives the image bytes.
Is this plugin needed with a native official vision model in DSH?
The README says no: with the official vision model, attach the image directly.
07Related DSH workflows
dsh-image-gen
by shanliuling
Adds in-chat multimodal image generation via a `generate_image` tool with multi-provider support (Gemini/OpenAI/Seedream), interactive preview, copy/download, and session-attached results.
dsh-vision-router
by ysr666
Eyes for text-only DeepSeek Harness agents: built-in free vision chain (no key) + pixel-level vision tools (Q&A, grounding, crop, pixel diff, colors, OCR, SVG trace, cutout, screenshots). One-command install, no Python, image turns work like ordinary tool-calling turns.
dsh-vision-toolkit
by anionex
为纯文本 DSH Agent 提供 10 个结构化视觉工具:意图感知图片问答、长截图 OCR、原始像素 grounding、UI 还原、像素 diff 等
dsh-see-now
by saos-3624
Captures screen, OBS, or webcam frames, filters by visual change, packs them into image sheets, and delivers them to DSH sessions.
08Data and sources
为 [DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness) 打造:GUI 附加图片自动转译,纯文本 DeepSeek 也能识图。
注册一条新提供商路由(`deepseek-vision`),包装真正的 DeepSeek 适配器
**如果你用官方视觉模型,直接发图即可,不需要本插件**。
## 许可证 MIT
This page is generated from the project’s public documentation, repository metadata and a structured parse of DSH Plugins; last verified on 2026-10-18. Found an error? Submit a correction.
Best DeepSeek Harness Plugins
Twelve plugins worth installing first — picked from the whole catalog, across every category.
