dsh-vision-analysis

A text-only agent cannot see images, and wiring up a vision model usually means juggling API keys, swapping models, and shuffling local files. dsh-vision-analysis removes all of that friction: install it, paste an image into the conversation, and get an analysis back — no key, no local file dance, no model switch. Out of the box it points at a free anonymous vision endpoint (OVHcloud AI Endpoints, Qwen2.5-VL-72B), so the very first call works with zero configuration.

The plugin ships eight tuned analysis modes — describe, OCR, chart-data, UI review, object detection, multi-image compare, code generation, and endpoint debug. OCR and chart-data return machine-readable JSON that downstream agents can consume directly, and the other modes each carry a purpose-built instruction template. It speaks both OpenAI chat/completions and Anthropic messages wire formats, so you can route through GPT-4o, Claude, Qwen-VL, a local Ollama instance, or anything in between. When one vision model hits its rate limit the plugin automatically fails over to the next in the chain and surfaces clear recovery guidance. Privacy is built in: image bytes are sent base64-embedded only to your configured endpoint, never enter the session log, and never reach the main model; API keys remain fully masked in every diagnostic report.

It is well suited for developers who want to give a text agent in DeepSeek Harness vision capability without standing up extra infrastructure, and for teams whose workflows involve bulk OCR, extracting structured data from charts, design review of interface screenshots, or generating HTML and CSS from a UI mock-up.