Routing pixels straight to vision models, not through text

What it does

The plugin intercepts image uploads in DeepSeek Harness chat and routes them to a dedicated vision model as ordinary tool calls, leaving the main DeepSeek text model untouched. It ships fourteen pixel-level tools—including vision_ground, vision_crop, vision_pixel_diff, OCR, SVG trace, and HTML screenshot—plus a built-in anonymous OVHcloud fallback that requires no API key. Agents can chain multiple vision steps in a single conversation turn, iterating on crops, diffs, and grounding coordinates until the task is done.

The interesting bit

Most DSH vision plugins transcribe images into text descriptions for DeepSeek, which discards spatial detail; this one keeps the raw pixels on the vision side and only feeds structured results back to the reasoning model. The built-in free tier is genuinely free—no account, no key, roughly 10 RPM across independent OVHcloud buckets—yet user-provided models still run first if configured.

Key highlights

• Pixel-faithful routing: original image data stays on the vision model instead of being flattened into a text caption
• Built-in anonymous fallback chain via OVHcloud endpoints; zero signup required out of the box
• Fourteen deep tools including pixel-diff verification, palette extraction, foreground cutout, and screenshot capture
• Caches vision responses by image content to avoid redundant backend calls
• Transparent UI handling: uploaded images still render normally in chat while the model call is internally rerouted

Caveats

• The anonymous free tier is rate-limited to roughly 2 requests per minute per IP per model (about 10 RPM in theory across buckets), so heavy use needs a personal vision key
• v1.5.3 specifically added guards against repetition-loop vision output and reworded misleading copy, suggesting earlier versions could spin on bad responses
• Requires Node.js 22 or newer and a DSH Web profile setup

Verdict

Worth a look if you run DeepSeek Harness and want agents that can inspect UI elements or verify pixel-level changes without swapping your entire session to a multimodal model. Skip it if you already have a dedicated vision API key and prefer a simpler description-based bridge.