Verified MCP server: Jina AI
The short answer
Last verified 2026-08-03
The Jina MCP server is Jina AI's own hosted server for reading web pages, searching the web and searching academic papers, and it answers on a web address instead of running as a program on your Mac. The repository behind it is jina-ai/MCP, Apache-2.0 licensed, 794 stars, and it was pushed the same day we captured this page. Twenty-one tools came back from the running server, two more than its README table lists. Several of them need no credential at all: we called primer with no Authorization header of any kind and got a real answer back, so the badge on this page rests on a call that actually succeeded. The rest go through paid Jina APIs, and we captured word for word what those say when the key is missing. All of it is below. [1][2][3]
Verification
Method
The server we verified is the one Jina AI hosts at https://mcp.jina.ai/v1, so there is no local process to spawn and no stderr to read. We spoke raw JSON-RPC over Streamable HTTP to it from this Mac, with node only as the test client: initialize, then notifications/initialized, then tools/list, then one real tool call. The server is stateless and issued no session id, so nothing had to be carried between requests. The tool call was primer, which needs no credential at all, and it answered with a genuine result derived from our own request. This Mac holds no Jina API key and none was used: the successful call went out with no Authorization header on it. We then called search_web twice, once with an obvious placeholder bearer token and once with no header at all, to capture what an auth-gated tool says when the key is missing, and ran two more probe passes covering a missing required argument, an unknown tool name, a wrong argument type and a malformed URL. The tool table below is the tools/list response, word for word. Two notes on the blocks that follow. The startup banner is the initialize reply with the long instructions string taken out so it fits on a page, and the untruncated capture is banked with the receipts. The sample call excerpt is the first four lines of what came back: the rest of that reply carried our own IP address, city-level location and network details, which we left out on purpose.
Startup banner
{"result":{"protocolVersion":"2025-06-18","capabilities":{"tools":{"listChanged":true}},"serverInfo":{"name":"Jina AI Official MCP Server","version":"1.4.0"}},"jsonrpc":"2.0","id":1}One real tool call
tools/call primer {}timestamp: utc: 2026-08-03T21:07:27.733Z userTimezone: Europe/Bucharest userLocalTime: 8/4/2026, 12:07:27 AM
Real lines from the captured result: enough to prove the call answered.
Tools
The server answered tools/list with 21 tools on 2026-08-03. The names, descriptions and parameters below are its own words, copied from that response and never edited.
| Tool | What it does |
|---|---|
| show_api_key | Return the bearer token from the Authorization header of the MCP settings, which is used to debug. |
| primer | Get up-to-date contextual information of the current session to provide localized, time-aware responses. Use this when you need to know the current time, user's location, or network environment to give more relevant and personalized information. |
| guess_datetime_urlurl* | Guess the last updated or published datetime of a web page. This tool examines HTTP headers, HTML metadata, Schema.org data, visible dates, JavaScript timestamps, HTML comments, Git information, RSS/Atom feeds, sitemaps, and international date formats to provide the most accurate update time with confidence scores. Returns the best guess timestamp and confidence level. |
| capture_screenshot_urlurl*firstScreenOnlyreturn_url | Capture high-quality screenshots of web pages in base64 encoded JPEG format. Use this tool when you need to visually inspect a website, take a snapshot for analysis, or show users what a webpage looks like. |
| read_urlurl*withAllLinkswithAllImages | Extract and convert web page content to clean, readable markdown format. Perfect for reading articles, documentation, blog posts, or any web content. Use this when you need to analyze text content from websites, bypass paywalls, or get structured data. |
| search_webquery*numtbslocationglhl | Search the entire web for current information, news, articles, and websites. Use this when you need up-to-date information, want to find specific websites, research topics, or get the latest news. Ideal for answering questions about recent events, finding resources, or discovering relevant content. |
| expand_queryquery* | Expand and rewrite search queries based on an up-to-date query expansion model. This tool takes an initial query and returns multiple expanded queries that can be used for more diversed and deeper searches. Useful for improving deep research results by searching broader and deeper. |
| search_arxivquery*numtbs | Search academic papers and preprints on arXiv repository. Perfect for finding research papers, scientific studies, technical papers, and academic literature. Use this when researching scientific topics, looking for papers by specific authors, or finding the latest research in fields like AI, physics, mathematics, computer science, etc. |
| search_ssrnquery*numtbs | Search academic papers and preprints on SSRN (Social Science Research Network). Perfect for finding research papers in social sciences, economics, law, finance, accounting, management, and humanities. Use this when researching social science topics, looking for working papers, or finding the latest research in business and economics fields. |
| search_jina_blogquery*numtbs | Search Jina AI news and blog posts at jina.ai/news for articles about AI, machine learning, neural search, embeddings, and Jina products. Use this to find official Jina documentation, tutorials, product announcements, and technical deep-dives. |
| search_imagesquery*return_urltbslocationglhl | Search for images across the web, similar to Google Images. Use this when you need to find photos, illustrations, diagrams, charts, logos, or any visual content. Perfect for finding images to illustrate concepts, locating specific pictures, or discovering visual resources. Images are returned by default as small base64-encoded JPEG images. |
| parallel_search_websearches*timeout | Run multiple web searches in parallel for comprehensive topic coverage and diverse perspectives. For best results, provide multiple search queries that explore different aspects of your topic. You can use expand_query to help generate diverse queries, or create them yourself. |
| parallel_search_arxivsearches*timeout | Run multiple arXiv searches in parallel for comprehensive research coverage and diverse academic angles. For best results, provide multiple search queries that explore different research angles and methodologies. You can use expand_query to help generate diverse queries, or create them yourself. |
| parallel_search_ssrnsearches*timeout | Run multiple SSRN searches in parallel for comprehensive social science research coverage and diverse academic angles. For best results, provide multiple search queries that explore different research angles and methodologies. You can use expand_query to help generate diverse queries, or create them yourself. |
| parallel_read_urlurls*timeout | Read multiple web pages in parallel to extract clean content efficiently. For best results, provide multiple URLs that you need to extract simultaneously. This is useful for comparing content across multiple sources or gathering information from multiple pages at once. |
| classify_texttexts*labels*model | Classify texts into user-defined labels using Jina embeddings. Use this when you need to categorize, tag, or sort text content into predefined categories. Perfect for sentiment analysis, topic classification, content moderation, or any text categorization task. |
| sort_by_relevancequery*documents*top_n | Rerank a list of documents by relevance to a query using Jina Reranker API. Use this when you have multiple documents and want to sort them by how well they match a specific query or topic. Perfect for document retrieval, content filtering, or finding the most relevant information from a collection. |
| deduplicate_stringsstrings*k | Get top-k semantically unique strings from a list using Jina embeddings and submodular optimization. Use this when you have many similar strings and want to select the most diverse subset that covers the semantic space. Perfect for removing duplicates, selecting representative samples, or finding diverse content. |
| deduplicate_imagesimages*k | Get top-k semantically unique images (URLs or base64-encoded) using Jina CLIP v2 embeddings and submodular optimization. Use this when you have many visually similar images and want the most diverse subset. |
| search_bibtexquery*numyearauthor | Search for academic papers and return BibTeX citations. Searches DBLP (computer science) and Semantic Scholar (broad academic coverage) for comprehensive results. Returns formatted BibTeX entries ready to use in LaTeX documents. |
| extract_pdfidurlmax_edgetype | Extract figures, tables, and equations from PDF documents using layout detection. Perfect for extracting visual elements from academic papers on arXiv or any PDF URL. Returns base64-encoded images of detected elements with metadata. |
Parameters marked with * are required.
Setup
Copy the block for the app you use. Each one is the configuration this server was verified with.
Open the file ~/Library/Application Support/Claude/claude_desktop_config.json (in Claude Desktop: Settings, then Developer, then Edit Config) and add:
{
"mcpServers": {
"jina": {
"command": "npx",
"args": [
"mcp-remote",
"https://mcp.jina.ai/v1",
"--header",
"Authorization: Bearer YOUR_JINA_API_KEY"
]
}
}
}That npx line is the mcp-remote bridge, which turns a hosted server into the local command this file expects. Newer Claude Desktop builds also take a plain url entry with a headers block, and the Connectors page takes the address on its own. If you have no key yet, delete the two --header lines: the server still connects and still lists every tool, and only the paid ones answer Unauthorized. Quit and reopen Claude Desktop after you save. [2][9][8]
One command in your terminal, and it is the one Jina documents:
claude mcp add -s user --transport http jina https://mcp.jina.ai/v1 --header "Authorization: Bearer ${JINA_API_KEY}"The -s user flag makes it available in every project instead of just this one. That command reads JINA_API_KEY from your shell, so export it first, or paste the key itself in place of ${JINA_API_KEY}. If you added this server earlier with --transport sse, remove it first with claude mcp remove -s user jina, then run the line above. [2][10]
Add to ~/.cursor/mcp.json for every project, or to .cursor/mcp.json inside one project:
{
"mcpServers": {
"jina": {
"url": "https://mcp.jina.ai/v1",
"headers": {
"Authorization": "Bearer YOUR_JINA_API_KEY"
}
}
}
}The url key is what marks the entry as remote, so there is no command to run and nothing to install. Cursor picks the file up on restart. Treat this file the way you treat any file with a key in it: it is plain text on disk, so keep it out of a repository you push. [11][2]
No JSON file and no terminal. In Routines: Settings, then Assistant, then Connections, then Add MCP Server. The form already opens in URL (SSE/HTTP) mode, which is the mode this server needs:
Name Jina AI Server URL https://mcp.jina.ai/v1 Bearer token YOUR_JINA_API_KEY
Leave the Command (stdio) toggle alone: this server is a web address, not a command. The Bearer token field is optional here, and an empty one still gets you the tools that need no key, primer and search_bibtex among them. Click Test Connection first, which opens the connection and reports the tool count without saving anything, twenty-one for this one. [12][2]
No terminal needed
If you have never opened Terminal and never want to, this is your path. Routines is a Mac app that connects MCP servers for you: paste one web address, and this server's tools are available to your AI in chat and in scheduled routines. The Jina key is the only decision to make, and the second step tells you whether you need one.
01
Download the app from getroutines.ai/download, drag it to Applications, and sign in.
02
Four tools need no credential at all: primer, guess_datetime_url, search_jina_blog and search_bibtex. Three more, read_url, parallel_read_url and capture_screenshot_url, work without one at a lower rate limit. Everything else, web search, arXiv and SSRN search, image search, reranking, classification and PDF extraction, runs against paid Jina APIs and needs a key from jina.ai. If you want those, create the key now and copy it.
03
Back in Routines, click your account at the bottom of the sidebar and choose Settings. Open the Assistant section, then the Connections tab, scroll to MCP Servers, and click Add MCP Server.
04
The form opens in URL mode already, which is what a hosted server needs, so leave the Command (stdio) toggle alone. Name: Jina AI. Server URL: https://mcp.jina.ai/v1. Keep the /v1 on the end.
05
Put your Jina key in the Bearer token field, or leave it empty to start with the free tools only. There is no Environment Variables box in URL mode: that field belongs to the Command (stdio) form, and this server does not use it. Click Test Connection: Routines opens the connection and reports how many tools it found, twenty-one for this one. Then click Add Server.
06
The tools work in chat right away, so you can paste a link and ask for the article back as clean text. To let a scheduled routine use them, open the routine, find the Tools & connections card, and tick the server under Apps.
Routine ideas
Once the server is connected, a scheduled routine can use its tools while you are away. Copy a prompt, paste it into Routines, and pick a time.
Every Monday at 8:00, search the web and arXiv for anything published in the last seven days on the topics listed in topics.md, read the five most useful pages in full, and save a one-line summary of each with its link to weekly-research.md.
Every weekday at 7:30, read the three competitor pricing pages listed in competitors.md, compare each one against what pricing-watch.md already says about it, and append a dated line for every change you find.
Every Friday at 16:00, search arXiv for new papers on the research topics in topics.md, fetch a BibTeX entry for each paper worth keeping, and append the entries to references.bib in my Notes folder.
Troubleshooting
Real errors captured during the verification run, printed exactly as the server returned them.
What you see
Error: Search failed for query "model context protocol": Unauthorized
The fix
This is the server working as designed. It hands every client all 21 tools whether or not a key is attached, and only checks the key when a tool that costs money actually runs. We got this exact line back twice on purpose, once with an obvious placeholder bearer token and once with no Authorization header at all. Check that the key is in the right place for your client: the Bearer token field in Routines, the headers block in Cursor, the --header flag in Claude Code, and the mcp-remote --header argument in Claude Desktop.
What you see
MCP error -32602: Input validation error: Invalid arguments for tool read_url: [
{
"code": "invalid_union",
"unionErrors": [
{
"issues": [
{
"code": "invalid_type",
"expected": "string",
"received": "undefined",
"path": [
"url"
],
"message": "Required"
}
],
"name": "ZodError"
},
{
"issues": [
{
"code": "invalid_type",
"expected": "array",
"received": "undefined",
"path": [
"url"
],
"message": "Required"
}
],
"name": "ZodError"
}
],
"path": [
"url"
],
"message": "Invalid input"
}
]The fix
The server validates arguments before it spends anything, so a malformed call never costs you credit. We triggered this by calling read_url with no url at all; the two branches in the message are the two shapes url accepts, one URL or an array of them. A wrong type comes back in the same -32602 shape: sending num to search_web as the string three produced Expected number, received string. Nothing is broken, the assistant simply has to send what the tool table above lists.
What you see
MCP error -32602: Tool read_urll not found
The fix
A plain typo, and we made this one on purpose by asking for read_urll. The tool table above is the full list the server answered with, so check the spelling against it. This is also what you see if a client cached an older tool list: disconnect the server and reconnect it so the client asks for tools/list again.
What you see
MCP error -32602: Input validation error: Invalid arguments for tool read_url: [
{
"validation": "url",
"code": "invalid_string",
"message": "Invalid url",
"path": [
"url"
]
}
]The fix
The address was not a full URL. We sent not-a-url to prove it. Give the tool the complete address including the scheme, so https://example.com/article rather than example.com/article. If you pasted the link out of a chat message, check that nothing was cut off the front of it.
The fix
This one never reaches Jina, so there is no server message to read. A hosted server needs a working internet connection at the moment the call runs, which matters for a scheduled routine that fires while you are on a plane. And a client that can only launch local commands reaches this address through the npx mcp-remote bridge, which belongs to Node.js: if Node is missing, that bridge fails at the spawn step before any request is sent. Install Node from nodejs.org, or use a client that dials the address directly, which is what Claude Code, Cursor and Routines all do.
FAQ
It is Jina AI's own hosted MCP server, published from the jina-ai/MCP repository and answering at https://mcp.jina.ai/v1. Once connected, an AI assistant such as Claude can search the web, read any URL back as clean markdown, screenshot a page, search arXiv and SSRN, fetch BibTeX citations, rerank or classify text, and pull figures out of a PDF. There is nothing to install, because the server runs on Jina's side and your client just calls the address. [1][2]
The running server answered tools/list with 21 tools: web work (read_url, parallel_read_url, capture_screenshot_url, guess_datetime_url), search (search_web, parallel_search_web, search_images, expand_query), research (search_arxiv, search_ssrn, parallel_search_arxiv, parallel_search_ssrn, search_bibtex, extract_pdf, search_jina_blog), embedding utilities (classify_text, sort_by_relevance, deduplicate_strings, deduplicate_images) and one debug helper, show_api_key, plus primer for the current time and location. The README table lists 19, so the live capture found two the documentation does not mention. The table above is that capture, word for word. [2]
Only for some of it. Four tools need no credential at all: primer, guess_datetime_url, search_jina_blog and search_bibtex. Three more, read_url, parallel_read_url and capture_screenshot_url, work without a key at a lower rate limit. The paid Jina APIs sit behind the rest, so web search, arXiv and SSRN search, image search, query expansion, reranking, classification, deduplication and PDF extraction all need a key from jina.ai. We proved both ends of that split live: primer answered with no header on the request, and search_web answered Unauthorized with no header and again with a fake one. [2][4]
Two things are worth knowing. This is a hosted tool, so every page you ask it to read and every query you ask it to run goes to Jina's servers by design, and your key travels there in an Authorization header on each call. And one tool, show_api_key, exists to print that bearer token straight back into the conversation, which is handy when a key is not being picked up and a bad idea in a transcript you are about to share. Nothing in the tool list writes to your Mac: every one of the 21 reads, searches, scores or reports something and hands the result back. [1]
No. jina-mcp-tools is a third-party package, maintained by PsychArch rather than by Jina AI, at version 1.2.5 with 256 downloads in the week ending 2026-08-02 and 44 stars on GitHub. It is the most popular of several unofficial wrappers around Jina's APIs, and some of the older ones are archived. The server on this page is Jina's own, and there is no npm package called jina-mcp-server: that name returns a 404 on the registry. [6][7][5][1]
Jina AI's own GitHub bio states the company was acquired by Elastic in October 2025, so it is fair to ask. The evidence on the day we checked points the other way from abandonment: the jina-ai/MCP repository was pushed on 2026-08-03, the same day we captured this page, it is not archived, and the live endpoint self-reported version 1.4.0. Worth rechecking before you build something long-lived on it, which is why the capture on this page carries a date. [3][1]
Any MCP client. Clients that dial a web address, such as Claude Code, Cursor and Routines, point straight at https://mcp.jina.ai/v1 and put the key in a header. Clients that can only launch a local command reach the same address through the npx mcp-remote bridge, which is the shape shown in the Claude Desktop block above. The server is identical either way; only the place you paste the config differs. [2][9][10][11][12]
Sources
Anything on this page we did not observe in the run is linked here, with the date we read it. The run itself is the receipt for the rest.
[1]
Described as the Official Jina AI Remote MCP Server: 794 stars, Apache-2.0, not archived, created 2025-08-09 and last pushed 2026-08-03, per the GitHub API.
[2]
The endpoint https://mcp.jina.ai/v1 with /sse kept as a legacy alias, the per-client install commands, and the table of which tools need a JINA_API_KEY. That table lists 19 tools; the live server answered with 21.
[3]
Org bio on the day of capture: "Your Search Foundation, Supercharged! (acquired by @elastic 2025/10)".
[4]
Where a Jina API key is issued. The separate rate-limit page at jina.ai/api-dashboard/rate-limit returned HTTP 404 on the same day.
[5]
404 Not Found: there is no npm package by that name, because this server is published as an endpoint rather than as a package.
[6]
The most popular unofficial alternative: version 1.2.5 published 2026-07-17, maintained by PsychArch, MIT, 44 stars on GitHub and last pushed 2026-07-29.
[7]
256 downloads for the week ending 2026-08-02.
[8]
The local bridge that lets a stdio-only client reach a hosted MCP server. MIT, maintained by threepointone.
[9]
The official Claude Desktop quickstart and the config file location.
[10]
The claude mcp add syntax for HTTP transports, scopes and headers.
[11]
The mcp.json shape, including url servers with a headers block, and the file locations.
[12]
How Routines runs one-click OAuth connectors and any MCP server.
This page describes Jina AI as it behaved in one dated run on one Mac. Versions move: if something here no longer matches what you see, the capture date at the top says how old the reading is.
Behind this directory
Routines, the app behind this directory, connects MCP servers like this one without a terminal: see how connectors work. The digests it writes stay markdown files on your Mac, and there is no cloud bill. Download Routines