MIT-licensed vision plugin for DeepSeek Harness: installed into the DSH Web profile with one dsh plugin command, it routes image turns to a free keyless vision chain and adds fourteen pixel-level tools — grounding, crop, pixel diff, OCR, SVG trace and more.
DSH integration
Native runtime
Author-claimed
Safety audit
Unaudited
Last verified
2026-08-21
License
MIT
01What can it help you accomplish?
Give text-only DeepSeek Harness agents pixel-level image understanding
Image Q&A answers, original-pixel grounding boxes, zoomed crops, OCR text, dominant colors, SVG traces and cutouts, all as ordinary tool calls
DSH users on text-only DeepSeek / opencode routes who need the coding agent to actually see pasted screenshots and design images
Verify UI rebuilds against a reference design with a measurable pixel loop
HTML screenshots plus per-pixel diff reports — diff ratio, red heatmap PNG and worst-region ranking — repeated until the mismatch converges
Frontend and coding-agent workflows restoring web UI from reference screenshots who want pixel-verified results instead of eyeballing
Run AI vision out of the box without API keys or paid accounts
A keyless anonymous OVHcloud fallback chain (five models, about 10 RPM in theory) with automatic classified failover, optionally fronted by user-configured or local backends
Hobbyists and teams who want image understanding in DSH immediately, before configuring any vision provider
02How to install into DeepSeek Harness
Prerequisites
- DeepSeek Harness Web profile (normal installs use `npx @deepseek-ai/dsh`; source checkouts use `pnpm dsh`)
- Node ≥ 22 on the host
- No API key for the default free chain; Chrome / Chromium / Edge only for `vision_html_screenshot`; Tesseract optional (OCR falls back to the vision model)
Installation steps
- 01
Run `npx @deepseek-ai/dsh plugin --profile web add dsh-vision-router` (from a DeepSeek Harness source checkout: `pnpm dsh plugin --profile web add dsh-vision-router`)
$ npx @deepseek-ai/dsh plugin --profile web add dsh-vision-router
- 02
Start or reload DSH Web; if installing into an already long-running Web process, let it reload once so the plugin bundle is discovered
- 03
Open the model selector in the lower-right corner of the chat composer and choose a group marked '+ Auto Vision'
- 04
Paste or upload an image normally — the agent can immediately use `vision_describe`, `vision_ground`, `vision_crop` and the other tools across multiple steps
Verify the integration
- `npx @deepseek-ai/dsh --profile web --dump-config | grep vision-router` (source checkout: `pnpm dsh --profile web --dump-config | grep vision-router`)
Rollback
- Run `npx @deepseek-ai/dsh plugin --profile web remove dsh-vision-router` — removes the dependency and the bundle layer; if you disabled the stock DeepSeek row manually, re-enable it in your profile patch
- Or set `disabled: true` for `vision-router` in the profile patch to unload wrapper routes, tools, skill and settings card (cached artifact files remain)
03DSH integration and capability boundaries
DSH Web profile plugin installed via the official dsh CLI; ships its own composition bundle patch and registers '+ Auto Vision' wrapper model groups inside DeepSeek Harness
Automatic vision routing ('+ Auto Vision' groups)
image turns in any DSH Web chat session→image turns are answered through a vision model while text turns keep using the daily model; original model groups are never modified
registers an admission wrapper and '+ Auto Vision' wrapper model groups in the DSH profileoptional stealth mode can take over the official deepseek-official route (off by default, explicit opt-in)Fourteen pixel-level vision tools
image attachments or local image files→grounding boxes, crops, pixel-diff heatmaps + JSON reports, OCR text, dominant colors, SVG traces, foreground cutouts and HTML screenshots — no Python required
artifacts are written under `<workspace>/.dsh-vision-router/artifacts``vision_screenshot` desktop capture is disabled by default and requires an explicit privacy opt-inMulti-backend fallback chain with keyless free endpoint
any vision tool call→backends tried in order — user vision models, optional local Ollama / LM Studio, custom HTTP endpoints, then the anonymous OVHcloud fallback — with classified errors and Retry-After-aware cooldowns
network calls to the configured vision provider hosts (optional proxy applies to vision hosts only; DeepSeek stays direct)the anonymous OVH endpoint is limited to 2 requests/minute per IP per modelImage memory (content-hash caching)
images discussed in the session→vision answers cached by attachment content hash; later text turns reuse the recorded description (marked as untrusted evidence) without re-spending vision calls
04Who is it for? When not to use it?
Good for
- DSH users on text-only DeepSeek / opencode routes who need the coding agent to actually see pasted screenshots and design images
- Frontend and coding-agent workflows restoring web UI from reference screenshots who want pixel-verified results instead of eyeballing
- Hobbyists and teams who want image understanding in DSH immediately, before configuring any vision provider
Not for
- The built-in free fallback is the anonymous OVHcloud endpoint, capped at 2 requests/minute per IP per model (about 10 RPM in theory across five models); heavier use needs one of the README's free key channels or a local backend.
- Requires a DeepSeek Harness Web profile and Node ≥ 22 on the host; Oh-DSH Desktop ≤ 0.1.5 needs plugin v1.4.2+ (earlier versions crash its bundled runtime), and Linux desktop capture needs ImageMagick `import` or `scrot`.
- Mixing legacy manual `cordis.patch.yml` plugin rows with `dsh plugin add` can register the plugin twice; upgrading from a pre-bundle-patch v0.x install requires deleting the old manual row or `dsh web` fails at startup.
05Compatibility, maintenance and safety notes
- The built-in free fallback is the anonymous OVHcloud endpoint, capped at 2 requests/minute per IP per model (about 10 RPM in theory across five models); heavier use needs one of the README's free key channels or a local backend.
- Requires a DeepSeek Harness Web profile and Node ≥ 22 on the host; Oh-DSH Desktop ≤ 0.1.5 needs plugin v1.4.2+ (earlier versions crash its bundled runtime), and Linux desktop capture needs ImageMagick `import` or `scrot`.
- Mixing legacy manual `cordis.patch.yml` plugin rows with `dsh plugin add` can register the plugin twice; upgrading from a pre-bundle-patch v0.x install requires deleting the old manual row or `dsh web` fails at startup.
MIT · actively maintained (latest release v1.7.3, 2026-08-20)
06Frequently asked questions
How do I install dsh-vision-router into DeepSeek Harness?
It is a native DSH plugin: run `npx @deepseek-ai/dsh plugin --profile web add dsh-vision-router` (from a source checkout: `pnpm dsh plugin --profile web add dsh-vision-router`), reload DSH Web, then choose a '+ Auto Vision' model group in the lower-right picker. The package ships its own bundle patch, so no manual file edits are needed.
Do I need an API key for vision?
No. The chain ends with a built-in anonymous OVHcloud fallback — no account, no key, 2 requests/minute per IP per model. You can put your own vision models, local Ollama or LM Studio ahead of it, or use any of the free key channels listed in the README for higher quotas.
Does it change my daily DeepSeek model?
No. The original model groups are never modified — the plugin creates same-name '+ Auto Vision' entries, and text turns stay untouched in model, cost and context. Stealth-mode takeover of the official route is off by default and requires explicit opt-in.
Where do my images and data go?
Uploads send only the selected image plus your question to the configured vision provider hosts (an optional proxy applies to vision hosts only; DeepSeek stays direct). Generated artifacts are written locally under `<workspace>/.dsh-vision-router/artifacts`, and image text is always treated as untrusted evidence.
What are the main limits?
The anonymous free fallback allows 2 requests/minute per IP per model (about 10 RPM across five models), the host needs Node ≥ 22, Chrome / Chromium / Edge is required only for `vision_html_screenshot`, and the desktop screenshot tool is disabled by default.
07Related DSH workflows
modlens
by liustack
Vision bridge for text-only models: paste an image, get structured JSON evidence (OCR, layout, semantics).
agent-vision-toolkit
by anionex
Vision toolkit and skills that give text-only LLMs eyes — multi-image understanding, image Q&A, OCR, frontend UI restoration and GUI automation, with optional agent integration.
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-image-gen
by shanliuling
AI image studio for DeepSeek Harness — generate, edit & compare images in chat, with 500+ prompts, gallery, multi-model workflows and ComfyUI.
08Data and sources
The package ships its own composition patch (`dsh.bundle.patch`): `dsh plugin add` wires the row, the admission wrapper…
npx @deepseek-ai/dsh plugin --profile web add dsh-vision-router
the plugin discovers the model groups enabled under **Settings → Models** and creates same-name auto-vision entries.
This page is generated from the project’s public documentation, repository metadata and a structured parse of DSH Plugins; last verified on 2026-08-21. Found an error? Submit a correction.
Best DeepSeek Harness Plugins
Twelve plugins worth installing first — picked from the whole catalog, across every category.