Chunking, embedding and reranking are three steps that usually live in three libraries, and prototyping a retrieval approach means writing glue before learning anything. Exposing all three as calls makes the experiment conversational — try a chunk size, embed, rerank, see whether the right document surfaces. The reading tools serve a different purpose and are the more commonly used half: a page converted to something a model can actually read is worth more than raw HTML. get_help shipping the documentation as a tool is a small kindness that keeps parameter questions out of a browser.
A client for Jina AI's API surface covering retrieval infrastructure and web reading: embeddings and reranking, URL and search-result reading in a form built for models, a deep research mode, plus segmentation and classification.
- embeddings and rerank are the two halves of a retrieval pipeline — one to build the index, one to order what it returns.
- read turns a single URL into an LLM-friendly version of the page, and search does the same across the results for a term.
- deepsearch combines searching, reading and reasoning for an investigation rather than a lookup.
- segment splits text into segments or counts its tokens, which is the chunking step every retrieval pipeline needs.
- classify does zero-shot classification without a trained model.
- get_help returns the full Jina documentation, so the parameters are readable in place.
JINA_API_KEY.
One command plus a key — npx -y jina-ai-mcp-server-nodejs, then supply credentials