Why Vision Plugins Matter
Most DSH plugins are text-only by design — the runtime passes text to text models. But real work involves screenshots, UI designs, and images. Vision plugins bridge that gap, letting your agent paste an image and get structured JSON back. This post compares the top three DSH vision plugins by capabilities and ease of use.
The Three Contenders
| Plugin | Stars | Core Feature | Multi-Image | Pricing |
|---|---|---|---|---|
| dsh-vision-router | 929 | Free built-in vision chain | Yes | Free |
| agent-vision-toolkit | 1,091 | Paste image → structured JSON | Yes | Free |
| modlens | 3,499 | First vision plugin for DSH | Limited | Free |
dsh-vision-router: The All-in-One
This plugin ships a built-in free vision chain — no API key required. It supports pixel-level vision tools including Q&A, grounding, crop, pixel diff, colors, OCR, SVG trace, and cutout. Installation is one command, no Python dependency. Best for users who want zero-config vision out of the box.
agent-vision-toolkit: The Power User Choice
Designed for text-only LLMs, this toolkit supports multi-image understanding, image Q&A, frontend UI还原, and GUI automation. It integrates with Codex, Claude Code, Pi, Oh My Pi, and OpenCode. The trade-off: slightly steeper learning curve, but unmatched flexibility for complex vision tasks.
modlens: The Pioneer
The first vision plugin for DSH, modlens established the pattern of pasting images to get structured JSON evidence. It covers OCR, layout analysis, and semantic extraction. While newer competitors offer more features, modlens remains the simplest option for basic image-to-text workflows.
Which One Should You Choose?
- For beginners: dsh-vision-router (zero config, free)
- For power users: agent-vision-toolkit (flexible, multi-agent support)
- For simple OCR needs: modlens (minimal setup)
Security Considerations
Vision plugins process images locally or via API. Check each plugin's privacy policy before uploading sensitive screenshots. None of the three plugins we tested ship dangerous install scripts — all three earned A-grade in our security scan.