Introduction

DeepSeek Harness (DSH) follows the “everything is a plugin” philosophy and aims to extend agent capabilities through modular components. When building text-based agents, directly acquiring and understanding image information is a common technical challenge. Arnoldkevin/prismrelay-mcp aims to solve this problem by using a local MCP server to give text models a “pair of eyes” so they can process visual input.

Plugin Overview

The plugin is named Arnoldkevin/prismrelay-mcp and is a vision-focused local MCP implementation. It does not include a vision model by itself; instead, it sends images to the external Agnes API for inference and returns the results to the text model. It is maintained by Arnoldkevin, and the code is released under the MIT license. Note that PrismRelay is an independent community project and is not officially affiliated with Agnes AI.

Core Capabilities

The plugin provides the following core capabilities:

  • Image Understanding: Supports Q&A about scenes and objects, screenshots, documents, charts, visible text, and image comparison.
  • Image Generation: Experimental feature for generating images with optional reference, final size, candidates, and moderation.
  • Image Editing and Composition: Experimental feature for editing and composition with input roles, retained prompts, and moderation.

The corresponding tools include:
* prismrelay_understand_image
* prismrelay_generate_image
* prismrelay_edit_image

Installation and Enablement

Before installation, prepare Node.js 18 or later and an API key for the Agnes platform.

export AGNES_API_KEY="your_api_key_here"
npx @deepseek-ai/dsh plugin --profile web add github:Arnoldkevin/prismrelay-mcp

Typical Use Cases

The plugin can be used in the following scenarios:

  1. Ask about what is visible in a photo and request evidence.
  2. Diagnose UI screenshots or error dialogs.
  3. Extract titles, totals, dates, or status labels from clean document images.
  4. Interpret trends and legends in charts.
  5. Compare two product images or two UI screenshots.
  6. Check whether a poster contains specified elements or visible text.

Applicable Scenarios and Notes

The plugin provides only local stdio MCP support and does not include remote hosting, shared services, OAuth, billing, or video generation capabilities.

  1. Privacy and Data Processing: Images are read by the local MCP process and sent to the configured Agnes API for inference; they are not fully processed on-device.
  2. License Limitations: The MIT license applies only to the code in the repository and does not grant rights to Agnes models, APIs, outputs, documentation, branding, or trademarks.
  3. Terms of Service: Agnes’s Terms of Service prohibit unauthorized resale, sublicensing, or providing services to third parties.
  4. DSH Compatibility: DeepSeek Harness is currently in developer preview and may require tracking upstream breaking changes.

Conclusion

This plugin provides text models with external visual invocation capabilities and is suitable for specific scenarios that require image input. For more details, see the project home page: https://github.com/Arnoldkevin/prismrelay-mcp