ysr666/dsh-vision-router
Free vision for text-only agents: built-in keyless vision chain plus pixel tools (Q&A, grounding, crop, pixel diff, colors, OCR, SVG trace, cutout, screenshots); paste an image to use it.
Install
via npm
Resolves a published tarball rather than cloning the repository, and installs without any extra setup. Swap `web` for your profile name if you run another one.
via GitHub · npm package
Installing from GitHub runs the project's build script, which pnpm blocks until you allowlist it — run the command once and pnpm prints the exact key to add under `allowBuilds` in ~/.dsh/profiles/web/pnpm-workspace.yaml.
What happened when we ran it
Installed cleanly when we ran it
Every command here is run in a throwaway container against a clean profile, and the result is whatever the harness recorded — not a guess from the source. Last run 11d ago.
Show it in your README
For maintainers: the badge serves this listing's latest sandbox verdict, so a re-run updates it on its own — and it links readers to the full result here.
Due diligence
What it is — dsh-vision-router provides free vision capabilities for text-only agents in DeepSeek Harness with pixel tools.
Who it is for — When your agent needs vision grounding or vision crop operations, this plugin is suitable. If your agent supports multimodal input natively instead of being text-only, this plugin is not needed.
Watch out — No obvious issues found. Sandbox test passed in new profile: the plugin was installed and registered by harness. The plugin executes shell commands but no issues occurred.
The verdict — I would install it because it supports continuous multi-step image processing until the task is complete.
Generated by grok-4.6, and a starting point rather than a verdict. Where it says a plugin installs or does not, that is from a real run in a clean profile — everything else is read off the repository. Trust the source over this.
dsh-vision-router is a DeepSeek Harness plugin that routes image turns to vision models while leaving DeepSeek responsible for reasoning. It keeps the original image available to the vision side instead of first converting it into a text description. The plugin uses agent/request waterfalls and the bundled dsh.bundle.patch to add a composition row, an admission wrapper and attachment limits. dsh plugin add applies that wiring automatically; taking over the official DeepSeek route is an optional stealth-mode setting, disabled by default.
Vision requests use user-provided vision models first, followed by a built-in keyless OVHcloud fallback chain. The README describes pixel tools for Q&A, grounding, cropping, pixel comparison, colour extraction, OCR, SVG tracing, cutout and screenshots. Named tools include vision_ground, vision_crop, vision_describe and vision_pixel_diff. Image answers are cached by image content, while uploaded images remain visible in the conversation UI and the routing rewrite stays inside the model call.
This is intended for text-only agents that need iterative image inspection without switching their main model. It is a poor fit where the existing harness already provides the required vision route, or where external anonymous endpoints are unsuitable: the fallback is rate-limited to 2 requests per minute per IP per model. The plugin also reaches the terminal surface because its local pipeline uses sharp, potrace, tesseract and system Chrome for image processing and HTML screenshots.
Image handling is performed as an ordinary tool-calling turn. A multi-step sequence can call vision_ground, then vision_crop, vision_describe or vision_pixel_diff, apply a change, and inspect a new screenshot. The uploaded image remains an image in the session UI; the internal rewrite that directs the model to vision tools is not added to the session log. Text turns continue using the normal model, cost and context path.
The README describes a pixel-tool set covering visual Q&A, grounding, cropping, pixel diffs, colour sampling, OCR, SVG tracing, cutout and screenshots. It explicitly shows these tool names:
| Tool | Documented use |
|---|---|
vision_ground |
Locate an element in an image |
vision_crop |
Crop an image region |
vision_describe |
Produce an image description |
vision_pixel_diff |
Compare image pixels |
User-provided vision models are tried first. If they are unavailable, the built-in OVHcloud anonymous fallback chain provides five models without an account or key. The README states a limit of 2 requests/minute per IP per model, with roughly 10 RPM in theory across independent buckets. Answers are cached by image content.
The package requires Node.js >=22 and does not require Python. Its processing pipeline uses sharp, potrace, tesseract and system Chrome; system Chrome is used for HTML screenshots. The README identifies dsh.bundle.patch as the composition patch applied by the plugin. A bundled DSH Web profile is also referenced by cordis.patch.yml.
Taking over the official DeepSeek route is available through an optional stealth-mode setting and is off by default. The README does not specify a configuration key or command-line flag for that setting.
Written from the project's own documentation and kept in sync with it. Where the two disagree, the source is authoritative — read the README on GitHub
Same category
zhu1090093659
A `describe_image` vision tool for text-only models: images (local path, URL, attachment) go to a configurable OpenAI-compatible vision endpoint and only the returned text enters the session.
Installed cleanly when we ran it
liustack
Vision bridge for text-only models: paste an image, get structured JSON evidence (OCR, layout, semantics).
Installed cleanly when we ran it
Anionex
为纯文本模型"看图“设计更好的视觉工具箱和技能，支持多图理解，图片问答，前端UI还原、GUI 自动化等，并可选无缝接入多个主流agent，直接识别粘贴图片｜ A vision toolkit and skill designed for text-only llms — image Q&A, long-screenshot OCR, frontend UI restoration, and GUI automation, with optional seamless integration for Codex, Claude Code, Pi, Oh My Pi, and OpenCode
Installed cleanly when we ran it
Anionex
Vision for text-only models: paste an image and the model switches to a Vision Toolkit variant for image Q&A, multi-image comparison, long-screenshot OCR, screenshot-to-UI reproduction, element grounding, and pixel diff. No API key by default — images are processed by the author-hosted free service, 100 per machine per day; configurable to your own provider.
Installed cleanly when we ran it
omdsh-dev
Chrome sidebar extension that lets DeepSeek Harness operate your browser directly, no vision capabilities required. 一款 Chrome 侧边栏扩展程序，可让 DeepSeek Harness 直接操控您的浏览器，无需视觉能力。
Installed cleanly when we ran it
oxbshw
DeepWatch, powered by Watch Skill: a local-first perception and verification stack for AI agents. Understand video, audio and screen activity, keep timestamped evidence, and independently verify what a tool or workflow actually did. MCP, CLI, REST, Web and DeepSeek Harness.
Installed cleanly when we ran it
