Every tool takes a `collection` name, which means the indexing work happens on the Jina side before this is useful at all — connect it to an existing collection rather than expecting it to ingest your data. Once that is true, cross-modal is the capability worth having: describing what you are looking for in words and getting images back is not something a keyword index can do.
A search client for Jina AI's neural search. Three tools query a named collection: by meaning over text, by visual similarity from an image URL, and across the two — text describing an image, or an image finding text.
- `semantic_search` takes a `query`, a `collection` and an optional `limit` (default 10)
- `image_search` takes an `imageUrl` and finds visually similar images in the same collection
- `cross_modal_search` takes a `mode` of `text2image` or `image2text`, so a sentence can find pictures or a picture can find text
- Every result carries an `id`, a numeric `score` and its `data`, so ranking is visible rather than implied
- Errors are typed and returned rather than thrown — invalid keys, bad parameters, rate limits, network failures and unknown collection names each come back with a code
Node.js 16 or higher, and a Jina AI account with an API key from `https://cloud.jina.ai/`. Published as jina-ai-mcp version 1.0.0; the documented path is clone, `npm install`, put `JINA_API_KEY` in a `.env` file, `npm run build`, then point your client at `build/index.js` with the key in the entry's `env`. Every call names a `collection`, so the collections have to exist on the Jina side first — this server searches them, it does not create or index them.