Best DeepSeek Harness Tool Plugins: Vision, Browser, Search
Nine curated DeepSeek Harness tool plugins — vision bridges, browser control, agent search, and a toolkit of deterministic utilities — ranked by stars.
DeepSeek Harness (dsh) agents get bash and file editing out of the box, but vision, browser control, and agent-tuned search are all plugin territory. The single highest-starred plugin in FindHarness's entire curated index — modlens, a vision bridge — lives in this category. Here are nine tool plugins from the tools-capabilities category, ranked by GitHub stars as of August 2026.
The picks at a glance
| Plugin | Stars | What it does | Link |
|---|---|---|---|
| modlens | 1,320 | Vision bridge: image → structured JSON evidence | Plugin |
| dsh-vision-toolkit | 337 | Image Q&A, long-screenshot OCR, UI reproduction, grounding | Plugin |
| notes | 140 | Export conversations as PNGs or Markdown notes | Plugin |
| dsh-browser | 96 | Chrome sidebar extension for direct browser control | Plugin |
| modsearch | 94 | Web/X search bridge with structured citations | Plugin |
| argo | 70 | Multilingual search across web, academic, code, finance | Plugin |
| dsh-vision-router | 61 | Free, keyless vision chain plus pixel tools | Plugin |
| dsh-computer-use | 18 | Accessibility-first macOS computer use | Plugin |
| dsh-toolkit | 15 | Ten deterministic tools, zero dependencies | Plugin |
All nine come from the tools-capabilities category in FindHarness's curated (source: awesome) index. None have been security-reviewed by FindHarness; read the plugin security checklist before installing.
Why vision bridges dominate this category
DeepSeek's own chat-completions route is text-only — the official docs state plainly that it "cannot be made to accept images through any configuration." If you're running dsh against the DeepSeek-native provider rather than a vision-capable model, that's a hard limit, not a settings problem. That single fact explains why three of the top five plugins in this category (modlens, dsh-vision-toolkit, dsh-vision-router) exist at all: they're bridges that let a text-only model reason about images by having the plugin do OCR/layout/semantic extraction first and hand back structured text instead of raw pixels.
The nine plugins
1. modlens — the vision bridge
github.com/liustack/modlens · 1,320 stars · MIT
The highest-starred plugin in FindHarness's entire curated index, not just this category. Paste an image and modlens returns structured JSON evidence — OCR text, layout information, and semantic description — rather than expecting the model itself to see the picture. This is the plugin already used as the install-command example in our guide to installing dsh plugins.
dsh plugin --profile web add github:liustack/modlens
2. dsh-vision-toolkit — broader vision tasks
github.com/anionex/dsh-vision-toolkit · 337 stars · MIT
A wider vision toolkit for text-only models: intent-aware image Q&A, long-screenshot OCR (for scrolling captures that don't fit one frame), UI reproduction, grounding (locating an element by description), and pixel diff. If modlens covers your case but you also need UI-specific tasks like reproducing a screenshot as code or diffing two renders, this is the broader option.
dsh plugin --profile web add github:anionex/dsh-vision-toolkit
3. notes — export conversations
github.com/zhaoolee/notes · 140 stars · npm: @zhaoolee/dsh-notes
Exports dsh conversations as Smartisan Notes-style PNGs, or creates and updates Markdown notes in a configured, account-scoped workspace. Useful if part of your workflow is turning a session's output into a shareable artifact rather than leaving it inside the Web UI. Published to npm, so this is a prebuilt install with no allowBuilds prompt.
dsh plugin --profile web add @zhaoolee/dsh-notes
4. dsh-browser — direct browser control
github.com/lum1104/dsh-browser · 96 stars · MIT
A Chrome sidebar extension that lets dsh operate your browser directly — the description specifically notes no vision capabilities required, meaning it's built around DOM-level control rather than screenshot interpretation. If your model can't see images at all (see the DeepSeek text-only note above) and you need browser automation anyway, this sidesteps the vision problem entirely rather than working around it.
dsh plugin --profile web add github:lum1104/dsh-browser
5. modsearch — structured web search
github.com/liustack/modsearch · 94 stars · MIT · npm: @liustack/modsearch
From the same author as modlens, and built on the same idea: a search bridge for text-only agents that returns structured JSON evidence (search results, fetched content, citations) rather than raw HTML or a plain-text summary the model has to parse itself. Note that dsh already ships official web search providers (web-search-exa, web-search-perplexity, web-search-deepseek) — modsearch is a community alternative, not a replacement for a missing capability, so compare both before picking one.
dsh plugin --profile web add @liustack/modsearch
6. argo — multi-domain search
github.com/taxueseek/argo · 70 stars · MIT
Search built specifically for agents, with multilingual coverage spanning web, academic, code, shopping, finance, news, and encyclopedia sources in one plugin. Broader in scope than modsearch, which is more narrowly a web/X search bridge — argo is the pick if your agent's tasks span multiple of those domains rather than general web lookups.
dsh plugin --profile web add github:taxueseek/argo
7. dsh-vision-router — free, keyless vision
github.com/ysr666/dsh-vision-router · 61 stars · MIT
Another vision bridge, distinguished by a "keyless" vision chain — meaning it doesn't require you to configure a separate vision-model API key to work — plus a set of pixel tools: Q&A, grounding, crop, pixel diff, colors, OCR, SVG trace, cutout, and screenshots. Paste an image to use it, same interaction pattern as modlens. Worth comparing directly against modlens and dsh-vision-toolkit if you're choosing one vision plugin rather than stacking several.
dsh plugin --profile web add github:ysr666/dsh-vision-router
8. dsh-computer-use — macOS computer use
github.com/anionex/dsh-computer-use · 18 stars · MIT
Accessibility-first macOS computer use, with a design emphasis on fresh observations, stale-state rejection (refusing to act on an outdated view of the screen), scoped permissions, and safe input handling. This is a materially higher-risk category of tool than the others on this list — it's controlling your actual desktop, not a sandboxed browser tab — so the "accessibility-first" and "scoped permissions" framing in the description is worth verifying against the plugin's actual permission model before granting it broad access.
dsh plugin --profile web add github:anionex/dsh-computer-use
9. dsh-toolkit — a deterministic utility belt
github.com/omdsh-dev/dsh-toolkit · 15 stars · MIT
Ten small, deterministic tools in one zero-dependency install: time, encoding, JSON, calculator, CSV, regex, markdown, diff, stat, and schema. None of these are AI capabilities in themselves — they're the kind of boring, reliable utility tools that save an agent from writing a throwaway script for a simple calculation or format conversion. A reasonable one to install early, given the small footprint and broad utility.
dsh plugin --profile web add github:omdsh-dev/dsh-toolkit
Picking a vision plugin: modlens vs dsh-vision-toolkit vs dsh-vision-router
All three solve the same underlying problem — a text-only model can't see images — but with different scopes:
- modlens is the most-adopted, general-purpose bridge: paste an image, get structured evidence back.
- dsh-vision-toolkit adds UI-specific tasks (reproduction, grounding, pixel diff) on top of the base OCR/Q&A pattern.
- dsh-vision-router emphasizes not needing a separate vision-model API key, plus a broader pixel-manipulation toolset (crop, cutout, SVG trace).
Installing more than one is possible but likely redundant — they overlap heavily in the base OCR/Q&A case. Start with modlens given its adoption, and add dsh-vision-toolkit only if you specifically need the UI-reproduction or grounding features it adds.
FAQ
Do I need a vision-capable model to use these plugins?
No — that's the point of a vision bridge. modlens, dsh-vision-toolkit, and dsh-vision-router are all built so a text-only model can work with images by having the plugin extract structured evidence first, rather than the model itself processing raw pixels.
Can DeepSeek's own models see images if I configure them correctly?
No. dsh's documentation states that DeepSeek's native chat-completions route is text-only and can't be made to accept images through any configuration — this is a routing-level limit, not a settings.yaml flag you can flip.
Is dsh-computer-use safe to install?
Read its permission model carefully before granting it broad access — computer-use tools operate your actual desktop, a materially different risk profile than a sandboxed file-edit or browser-extension tool. The description emphasizes "scoped permissions" and "stale-state rejection," but verify those claims against the plugin's actual code or documentation rather than taking the description at face value.
Why do modlens and modsearch share an author?
Both are from liustack and follow the same design pattern: bridge a capability the base model lacks (vision, live search) by having the plugin return structured JSON evidence instead of raw content. If you like one, the other is worth evaluating with the same expectations.
Does dsh have any of these capabilities built in?
Web search, yes — dsh ships official search providers (web-search-exa, web-search-perplexity, web-search-deepseek, plus a generic HTTP fetch provider). Vision and browser/computer control are not built in; they're entirely plugin territory, which is why this category exists.
Next steps
For workflow and multi-agent orchestration plugins that often pair with these tools, see Best DeepSeek Harness Workflow and Automation Plugins. Browse the full category at /categories/tools-capabilities, or see our original pick list in 10 Best DeepSeek Harness Plugins in 2026.