dsh-multimodal-router

Running DeepSeek locally gives you strong reasoning, but the model still cannot see a single pixel. Paste a UI screenshot or a data chart and you get a polite refusal. dsh-multimodal-router closes that gap: it quietly delegates image perception to a vision-capable model on your local Ollama instance, converts the result into text evidence, and hands it back to DeepSeek — no manual prompt engineering, no endpoint configuration, no model name to remember.

Three pillars define the plugin. First, zero-config discovery: it scans Ollama model metadata to find Gemma 4 and other vision-ready models, accepts images via drag-and-drop, clipboard paste, or a paperclip button, and ships presets for OCR, charts, UI layout, and visible code. Second, local realtime voice: when an audio-capable model (Gemma 4 E2B, E4B, or 12B) and Qwen3-TTS are both available, the composer unlocks a full voice loop — incremental transcription every 1.5 seconds, silence-to-send, local TTS playback, and click-to-interrupt. Third, privacy by default: all routing stays on localhost, remote fallback is off unless you explicitly enable it, API keys are read only from environment variables, and archived filenames carry just a SHA-256 digest with no original name.

If you are a local-LLM developer on Apple Silicon or x86 Mac who wants multimodal capability without images ever leaving the machine, or simply curious about a fully offline realtime voice conversation, this plugin is a low-friction path: install it, drop in an image, and go. Voice components are optional and never break the image pipeline.