dsh-web-search-doubao

The built-in web search in DeepSeek Harness rides on the model inference pipeline—every search burns token quota and blurs into your Ark/LLM invoice. dsh-web-search-doubao routes searches through ByteDance's standalone Doubao Search API (Volcengine), completely decoupling retrieval from model inference. No Doubao chat model is involved. Every account gets 500 free searches per month; beyond that, pay-per-use or monthly packages apply, and none of it touches your model token bill.

The plugin registers into the ctx.web capability seam exactly like the built-in DeepSeek provider, so dsh-tool-web and everything above the seam work unchanged. Two editions are available: custom honors the requested result count server-side, while global returns longer structured document bodies. API key, Base URL, and Edition are all runtime-configurable through a dedicated card in the Harness settings GUI—save and the next search picks up the change, no restart needed. The index draws on ByteDance-ecosystem Chinese content (Toutiao, Douyin, etc.), with publish timestamps down to the second when a source provides one. The plugin maps only fields the API actually returns, never requests images, and stores credentials in the credentials service rather than any config file.

Built for DeepSeek Harness developers who want to separate search spend from model spend and prefer strong coverage of the Chinese web. If you lean toward Zhipu/GLM, the author maintains a sibling plugin—dsh-web-search-zai—sharing the same design language.