TZHR-invest/dsh-plugins#dsh-vision-tool
Agent-callable vision tool that describes local images via any OpenAI-compatible vision endpoint you configure, with an optional multi-model cross-check and no built-in keys.
An agent-callable vision tool for the DeepSeek Harness: it describes local images on demand via any OpenAI-compatible multimodal model, so text-only models can 'see' files in the workspace by handing their paths to this tool.
Install
dsh plugin --profile web add dsh-vision-tooldsh plugin --profile web add dsh-vision-tool (npm dsh-vision-tool 0.1.1, verified npm 2026-08-31; part of the TZHR-invest/dsh-plugins monorepo).
Compatibility
DSH web profile; uses any OpenAI-compatible multimodal model endpoint configured by the user.
Details
- Repo: TZHR-invest/dsh-plugins#dsh-vision-tool
- Category: Vision & Multimodal
- Stars: 2
- Version: npm dsh-vision-tool 0.1.1
- Last push: 2026-08-17
- First seen: 2026-08-14
Recent updates
Agent-callable vision tool; describes local images via OpenAI-compatible multimodal endpoint; path-based workflow.
FAQ
- Which model does it use?
- Any OpenAI-compatible multimodal model you configure — the endpoint and key are user-supplied, so it works with qwen-vl, gpt-4o, GLM-4V, and similar providers.
- Can text-only models use it?
- Yes — that is the point: the tool is agent-callable, so a text-only model can invoke it with a local image path and receive a description.
- Does it upload images anywhere?
- Images are sent to the configured multimodal endpoint for analysis; no other egress is involved.
Alternatives
RRRosmontis/dsh-qwen-mm · Yuuz12/dsh-vision-helper · 54xkeee/dsh-youreyes