Overview¶
DeepSeek Harness (DSH) does not have vision capability by default. This plugin allows integration with existing local vision models, enabling the assistant to recognize image content. Data is sent only to local models and does not pass through the cloud, making it suitable for privacy-conscious scenarios.
Core Features¶
- local_vision tool: Reads local images, calls an OpenAI-compatible endpoint for analysis, and returns a text description.
- Graphical settings interface: No need to modify configuration files; directly enter the endpoint URL, model name, and API Key on the settings page.
- Data stays on your machine: Images are sent only to locally running services, ensuring privacy and security.
- OpenAI-compatible endpoint: Supports the standard interface and can connect to local services such as LM Studio, Ollama, and vLLM.
Installation and Enablement¶
Install using the official command:
dsh plugin --profile web add "github:TIPSONG/dsh-vision-api-localorweb#main"
After installation, restart the Web client to apply the changes.
Configuration¶
Open Settings → Vision Model API Connectivity, and enter the following information:
| Field | Description | Example |
|---|---|---|
| Endpoint | OpenAI-compatible base URL | http://localhost:1234/v1 |
| Model Name | Locally loaded model ID | gemma-4-e4b |
| API Key | Optional authentication header; leave empty for local services | (empty) |
If the endpoint is left empty, the vision feature will be disabled.
Usage¶
- Enable vision: After configuring the endpoint and model name, save the settings.
- Provide an image: In the conversation, tell the assistant the image path, for example:
Take a look at this image:
C:\Users\me\Desktop\photo.png, and describe its content - Retrieve results: The assistant will call the
local_visiontool to read the image and return descriptive text.
Notes¶
- Image sending limitation: Sending images by dragging or pasting may be rejected. It is recommended to use an image path for vision.
- Privacy and security: Images are only sent to your local model and are not uploaded to the cloud.
- Compatible models: Ensure that the local service provides an OpenAI-compatible
/v1/chat/completionsendpoint and supports image input.
Ecosystem Background¶
DSH uses a plugin-based architecture. This plugin is contributed by the community and has no affiliation with the official DeepSeek team. For more details, visit GitHub or the Community Catalog.