dsh-vision

Eyes for a text-only DeepSeek. Registers a view_image tool that forwards the model's question about an image to any OpenAI-compatible vision endpoint and returns the answer as text. The default model (Zhipu glm-4.6v-flash) is free; DashScope, Volcengine, Moonshot, local Ollama, and future DeepSeek vision APIs work with the same configuration surface.

Features

• view_image tool — accepts an absolute local path, an http(s) URL, or a data: URL; answers arbitrary questions (OCR, counting, chart reading, layout).
• Model guidance — a system-prompt section teaches the model to call the tool whenever an image matters, with one focused question per call.
• Backend fallback — on 429/404/5xx, retries the next model in fallbackModels (defaults to the free-tier chain on the default endpoint).
• Thinking-model support — strips inline … reasoning; reasoning-only answers fail with actionable guidance.
• Safe by default — API key is never logged (redacted from error messages), requests carry a hard timeout, responses are capped at 2 MB, and local files are size-limited before base64 upload.

Install

Download the prebuilt bundle (recommended)

1. Get dsh-vision-0.1.0.zip from Releases and verify it against SHA256SUMS.txt.
2. Extract anywhere and double-click the installer: install.bat (Windows) or install.command (macOS/Linux).
3. Restart the dsh web process and hard-refresh the browser.

Uninstall with uninstall.bat / uninstall.command.

From source

Requires Node.js ≥ 20.

git clone https://github.com/xzyonline/dsh-vision.git
cd dsh-vision
npm install
node scripts/install.mjs

The installer builds lib/, links the package into the shared profile directory (, so every profile resolves it), and appends one row to (backed up first). It is idempotent. See for per-platform details.