dsh-doubao-plugin

In DeepSeek Harness, models lack built-in web search, image generation, and multimodal image recognition by default, and users typically have to configure MCP servers or third-party tools to accomplish these tasks. dsh-doubao-plugin registers a dynamic tool called doubao_ask through the local Doubao Relay OpenAI-compatible endpoint, allowing the model to invoke Doubao directly in conversation for search, image generation, and image understanding with no external dependencies.

The plugin covers three core operations: real-time web search; image generation with results displayed inline as Markdown; and multimodal image recognition that accepts URLs, Base64 data URLs, or local file paths, automatically compressing local images over 512 KB to 800 px width before sending. The paste-to-path mechanism intercepts image paste events in the input box for text-only models, saves the image to a local temporary path, and inserts that path as plain text, so even text-only models can trigger Doubao image recognition indirectly. Generated images are deduplicated, collapsing CDN mirrors and template variants into a single link, and returned as a structured images array.

Ideal for users who already run Doubao as a backend model in DSH and want to complete search, generation, and recognition workflows through natural-language conversation alone. Once the local Doubao Relay is running, installing the plugin and restarting Harness is all it takes for the model to discover and use the doubao_ask tool, turning multi-step multimodal tasks into a single message.