dsh-web-search-litellm

If your DeepSeek Harness web_search keeps failing with an Authentication Fails error because the API Key on your chat model page is actually a LiteLLM proxy credential, or if your company requires all model traffic to flow through a LiteLLM gateway and blocks direct access to api.deepseek.com, dsh-web-search-litellm is built for that exact situation. It routes the web_search call through the OpenAI Responses protocol to your LiteLLM proxy, where the DeepSeek Responses API executes the search natively on the server side and returns a grounded answer together with the real source URLs the model opened. No new API key is needed, and no third-party search service such as Tavily, Brave, or Exa is introduced—the search stays on DeepSeek's official infrastructure and is billed through your existing LiteLLM route.

After installation, you simply confirm the baseURL, model, and apiKeyEnv values in the web-search-litellm section of the Settings UI, and web_search is fully wired into the proxy chain. The plugin writes no custom session events, so older Harness builds load sessions without issue; the search process remains fully visible through the standard web_search tool call and result events.

It is a good fit for teams where all company traffic must traverse a LiteLLM proxy, individual developers whose DEEPSEEK_API_KEY actually holds a proxy key and causes authentication errors, integrations that prefer the OpenAI Responses protocol over the Anthropic /messages format, and anyone who wants to eliminate free-tier or third-party search dependencies and keep search execution on the official server side.