william-jin-cmu/dsh-vision
dsh plug-in: add vision to plain text DeepSeek - view_image tool bridges any OpenAI compatible VLM (default Wisdom Spectrum free file, measured 4 manufacturers and 10 models)
dsh-vision adds a view_image tool to text-only DeepSeek models and forwards the image plus the model's question to a configurable OpenAI-compatible VLM. That makes OCR, chart reading, screenshot diagnosis, object counting, and interface inspection available from Web, TUI, or remote sessions without moving the whole conversation to a multimodal model. It supports hosted GLM, Qwen and Doubao endpoints as well as local Ollama, strips leaked reasoning blocks from thinking models, and includes a fallback chain for capacity-limited free GLM models.
Install
git clone https://github.com/dsh-external/dsh-vision ~/dsh-plugins/dsh-visionThis is only step 1 of the README's verified native-mount path. The current pnpm layout also requires symlinking @deepseek-ai/dsh-tools and schemastery from the DSH checkout, then adding the plugin's lib/index.js to ~/.dsh/config.yaml. DSH Companion already bundles the plugin. Follow the repository README rather than treating this clone alone as a complete install.
Compatibility
All DSH entry points after host registration: Web, TUI, and remote channels. Sends images to an OpenAI-compatible vision endpoint; local Ollama endpoints need no API key. Current native setup depends on DSH checkout paths and pnpm symlinks.
Details
- Repo: william-jin-cmu/dsh-vision
- Category: uncategorized
- Stars: 33
- Version: latest from GitHub default branch
- Last push: 2026-08-13
- First seen: 2026-08-05
Recent updates
The current README documents a free GLM fallback chain, OpenAI-compatible backend switching, environment-variable credential precedence, cancellation propagation, and tested latency across multiple VLM families.
FAQ
- Is the clone command a complete dsh-vision installation?
- No. The verified native path also links two host dependencies and adds lib/index.js to ~/.dsh/config.yaml. Follow all three README steps.
- Which vision providers can dsh-vision use?
- Any compatible /chat/completions vision endpoint, including documented GLM, DashScope Qwen, Doubao Ark, and local Ollama configurations.
- Does dsh-vision require a paid API?
- No. The documented default uses a free GLM vision model, and a local Ollama OpenAI-compatible endpoint can run without an API key.
Alternatives
oil-oil/dsh-vision · liustack/modlens · Anionex/dsh-vision-toolkit