dsh-vision-mcp

August 14, 2026 · View on GitHub

中文readme参考README.cn.md

A first-party plugin for DeepSeek Harness that provides a describe_image tool: send an image file (PNG/JPEG/WebP/GIF) to a Qwen-VL vision model and get back a text description of the image content.

Zero-intrusion: registered through the standard DSH tool DSL (defineTool) — the model can call it as soon as the plugin loads.
Secret-safe: the API key is resolved per call through the DSH credential seam (ctx.credentials.resolve('QWEN_VISION_API_KEY')); it never lives in a config file, and rotating the key needs no restart.
Hot-reload: mounted via cordis.patch.yml (the user patch layer); DSH's config-level HMR applies it immediately without restarting the host.

Compatibility

Developed and verified against DeepSeek Harness 0.1.0-rc.5 (commit 47f9438).

DSH is in developer preview and ships compatibility-breaking changes (see the root README). This plugin only uses two long-lived extension points — the tool DSL (defineTool / ctx.tools.register) and the credential seam (ctx.credentials.resolve) — which the entire plugin ecosystem depends on, so it is expected to keep working across most updates. After a DSH upgrade, verify: the describe_image tool still appears in the model's catalog and answers successfully.

Installation

Option 1: Let your DSH guide you (recommended)

Send the following to your DSH and it will walk you through the install:

Please refer to this file https://github.com/CeasarSmj/dsh-vision-mcp/blob/master/docs/INSTALL.md to install the plugin

Your DSH follows docs/INSTALL.md automatically: configure the credential → install the package → mount the config → verify the tool.

Option 2: Official channel (`dsh plugin` CLI)

On any machine with the DSH CLI:

# Install from GitHub (registers the bundle and appends it to the profile's bundles list)
dsh plugin --profile <profile-name> add github:CeasarSmj/dsh-vision-mcp

# Or from npm (once published)
dsh plugin --profile <profile-name> add dsh-vision-mcp

The bundle's cordis.patch.yml inserts the plugin row automatically; it takes effect in new profiles or after a restart.

Option 3: Manual (hot-reload on a running profile)

Put this package into the profile's node_modules (C:\Users\Administrator\.dsh\profiles\node_modules\dsh-vision-mcp — a real directory or a junction both work);
Append to the profile's cordis.patch.yml:

- insert:
    - id: dsh-vision-mcp
      name: dsh-vision-mcp
      config:
        baseUrl: https://dashscope.aliyuncs.com/compatible-mode/v1
        model: qwen-vl-max

DSH's HMR loads the plugin on save — no restart needed.

Configuration

Item	Where	Description
API key	DSH credential store (`$DSH_HOME/.credentials.yaml`) or Settings → Credentials	Key name `QWEN_VISION_API_KEY`
`baseUrl`	plugin row `config.baseUrl`	OpenAI-compatible base URL; defaults to `https://dashscope.aliyuncs.com/compatible-mode/v1`. Point it at your own gateway when using Alibaba Cloud Bailian MaaS
`model`	plugin row `config.model`	Defaults to `qwen-vl-max` (`qwen-vl-plus`, `qwen2.5-vl-72b-instruct`, … also work)

Credential example ($DSH_HOME/.credentials.yaml):

QWEN_VISION_API_KEY: sk-xxxxxxxx

Tool: `describe_image`

file_path (required): path of the image (absolute, or relative to the harness working directory)
prompt (optional): what to focus on (e.g. "extract the text in this image", "describe the UI layout"); defaults to a detailed description

Once the model sees the tool, any task that needs to look at an image (screenshots, charts, photos, document pages) triggers it automatically.

Development

git clone https://github.com/CeasarSmj/dsh-vision-mcp
cd dsh-vision-mcp
# No build step: plain ESM, loaded directly by the DSH loader

index.js — the plugin itself (defineTool registration + Qwen-VL call)
cordis.patch.yml — the bundle config layer
docs/INSTALL.md — the conversation-guided installation doc

License

MIT