dsh-vision-router gives text-only DeepSeek Harness agents 14 image tools

dsh-vision-router, a plugin that gives text-only DeepSeek Harness agents the ability to work with images, has reached 1,083 GitHub stars since developer ysr666 created the repository on 13 August. The counts come from the GitHub REST API, retrieved 7 September: 1,083 stars and 47 forks in 25 days, MIT licensed, currently at version 2.1.0 on Node 22 or newer.

The plugin sends the image turn to a vision model with the pixels intact. The README contrasts this with most DSH vision plugins, which convert an image into a text description before the session sees it, losing the pixels on the way. DeepSeek stays the reasoning model for every text turn, and the vision model is called on demand, so looking at a picture becomes an ordinary tool call inside the session, with answers cached by image content.

Fourteen tools are built on that routing: visual question answering, grounding, cropping, pixel diff, colour and palette extraction, OCR, SVG tracing, cutout, description and HTML screenshots among them. The processing pipeline runs on sharp, potrace, tesseract and the system's Chrome, with no Python requirement. That toolset supports a loop the README demonstrates: locate an element, crop it, compare the pixels against the expected state, fix the code, screenshot again, and repeat until the work is correct.

The free tier is the unusual part. Vision tools fall back to a chain of five models on OVHcloud's anonymous endpoint, with no account and no key, limited to 2 requests per minute per IP per model. Vision models the user configures run first, and the free chain answers only when no key is set. The current 2.1.0 release adds a native settings panel, an explicit Vision mode toggle in the composer, runtime language switching, and sets the minimum supported harness version at the 0.1.2 release candidate, which ties the plugin to a preview version of the harness, as with the other DSH plugins.

For a team running agents on a text-only model, the capability is self-verification of visual work. An agent that edits a UI can now check what actually rendered, at the pixel level, without the whole session moving to a more expensive multimodal model: the text model keeps its cost and context, and the vision model is billed only on the turns that need to see an image. The rate-limited free chain is enough to evaluate the plugin and no more, so adopting it for real work means bringing a vision model key. The star count, 1,083 stars in 25 days, is the evidence that text-only harness users consider the absence of image support a real gap.

Source

Primary source: ysr666/dsh-vision-router on GitHub. Star and fork counts from the GitHub REST API, retrieved 2026-09-07.