dsh-gemini-multimodal

DeepSeek models are text-only by design, so a screenshot, a voice memo, a PDF, or a video clip is effectively invisible to them. dsh-gemini-multimodal closes that gap by delegating perception to Gemini while keeping all reasoning and decision-making in the main DeepSeek model, giving the Harness genuine multimodal awareness without changing its core identity.

The plugin ships four tools: media_understand handles images, audio, video, or URLs for OCR, chart reading, UI analysis, and spoken-word extraction; media_transcribe produces verbatim transcripts with timestamps where useful; image_generate turns a text prompt into a locally saved PNG; and read_document summarizes or answers questions about PDFs, Office files, and plain text. Two providers are available: the Gemini REST API for a single fast HTTP call, or the local Antigravity CLI signed in with a Google account for zero-configuration use.

If your DeepSeek Harness workflow regularly involves design mockups, meeting recordings, product demos, or client-delivered documents, this plugin removes the friction of manually describing media and pasting the text back in. The agent simply sees and hears the raw material, and you get the answer you actually need.