Free trial, then Pro from $6.67 a month.

Try it free

Guide12 min read

A Jina AI Reader alternative: when imread fits, and when it doesn't

nglsongChecked against the code as it ships
On this page11 sections
  1. 1The short answer
  2. 2What Jina Reader is, and why people look for alternatives
  3. 3imread in one minute
  4. 4Switching a script from r.jina.ai to imread
  5. 5What we measured on real pages
  6. 6Where imread is the better fit
  7. 7Where Jina Reader is the better choice
  8. 8Other routes besides a hosted reader
  9. 9Questions people ask
  10. 10How to choose, step by step
  11. 11How we checked this, and what we did not check

imread is a Jina AI alternative for one job: turning article-like web pages into clean text for an LLM. It answers the same address-prefix call as Jina Reader, needs no API key, and reads WeChat Official Account articles, GBK and Shift_JIS pages and SEC filings. It does not run JavaScript, take screenshots, search the web or accept uploaded files. If you need any of those, Jina Reader is the right tool and this guide will say so plainly.

This is a practical comparison for developers who already call r.jina.ai or are about to. It covers what the two services have in common, how to switch a script in a few minutes, what we measured on real pages on November 12, 2026, which of Jina's options have no counterpart in imread, and how to decide. The imread numbers below were measured by us. Everything we say about Jina Reader comes from its public README, and we did not run Jina's hosted service for this article.

The short answer

  • Choose imread when the pages are articles, documentation, news, blog posts, filings or WeChat posts, when you want no key and no sign-up, and when plain, predictable output matters more than rendering options.
  • Choose Jina Reader when pages are built in the browser with JavaScript, when you need web search, screenshots, raw HTML, image captions, uploaded PDF or Office files, or when you want to host the service yourself.
  • Use both if your pipeline mixes the two kinds of page. The call is the same shape, so routing a URL to one or the other is a one-line decision.

What Jina Reader is, and why people look for alternatives

Jina Reader is a service from Jina AI. You put https://r.jina.ai/ in front of a page address and get back text that is friendlier to a language model than raw HTML. A second host, s.jina.ai, takes a search query, searches the web, fetches the top five results and returns each as text. Its README describes a hosted API, an open-source branch under the Apache-2.0 license, and a Docker image that bundles headless Chrome, LibreOffice and CJK fonts.

It is a capable product, and most reasons to look elsewhere are about fit rather than quality:

  • Keys and quotas. Jina's own troubleshooting section says anonymous traffic is the most aggressively rate-limited, and that an API key gives a higher quota and access to features such as its internal proxy. If you want to call a reader from a script, a notebook or an agent without managing a key, that is friction.
  • A large option surface. The README lists dozens of request headers: engines, response timing, selectors, link and image handling, chunking, presets and proxies. That power is useful when you need it. When you only want the article, it is a lot to learn and a lot that can change an answer.
  • Output that varies with the options. A browser engine, a wait timing and a cache lifetime all influence what comes back. Some teams prefer a reader that does one thing the same way every time.
  • Where a page is sent. Any hosted reader sees the addresses you send it. Some teams want to know exactly what the service logs.
  • Specific content. Pages in older Chinese or Japanese encodings, WeChat articles and SEC filings are common sources for research agents, and a reader tuned for them can save a debugging session.

imread in one minute

imread is a reader for web articles. In a browser you paste a link and read the page without menus or ads, listen to it, translate it or copy it as Markdown. For developers and agents it offers a REST API, an MCP server and, since this month, the address-prefix call that Jina Reader made familiar:

curl "https://imread.org/https://paulgraham.com/greatwork.html"

The answer is text/plain in the same shape Jina's default response uses: a Title: line, a URL Source: line, then Markdown Content: followed by the article. Send Accept: application/json and you get Jina's JSON envelope instead:

curl -H "Accept: application/json" "https://imread.org/https://example.com"

{ "code": 200, "status": 20000,
  "data": { "title": "Example Domain", "description": "…",
            "url": "https://example.com/", "content": "…" } }

Send X-Respond-With: text to get plain paragraphs instead of Markdown. markdown and default are the same as sending nothing. The page's own query string stays part of its address, so https://imread.org/https://example.com/search?q=a&page=2 reads that exact page. Errors come back as one readable sentence, or as a JSON object with the same code and status fields when you ask for JSON.

The same extraction is also available at /api/extract with a url parameter, and through the MCP server. All three share one limit of 30 pages a minute per caller without a key. Pages are not stored: each request fetches the page afresh, and a read writes one anonymous log line that names the kind of page and how it went, not the address.

Switching a script from r.jina.ai to imread

For the common case of reading an article, the change is the host. Replace https://r.jina.ai/ with https://imread.org/ and remove the Authorization header, because imread needs no key (an optional Pro key only raises the limit). If your code reads the plain-text answer, the Title: and URL Source: lines are in the same place. If it reads JSON, the fields under data are the same four: title, description, url and content.

What does not carry over is Jina's other request headers. imread's prefix call reads two headers only, Accept and X-Respond-With, and ignores the rest. It never fails because you sent a Jina header it does not know, but it also will not do what the header asks. This table lists the Jina options a script is most likely to use:

Jina Reader headerWhat it does in JinaIn imread
x-respond-with: markdownMarkdown without the readability filterAccepted; the answer is always the main article as Markdown
x-respond-with: textPage text as plain textAccepted: plain paragraphs of the main article
x-respond-with: html, screenshot, pageshot, frontmatterRaw HTML, screenshot URLs, Markdown with a YAML headerHTTP 400 with an explanation
x-engine, x-wait-for-selector, x-timeoutPick headless Chrome or curl, wait for an element, wait longerIgnored; imread fetches the HTML once, with a 9 second budget
x-target-selectorReturn only the element a CSS selector matchesIgnored; imread picks the main article itself
x-max-tokens, x-token-budgetTruncate to a token count, or reject over a budgetIgnored on the prefix call; use offset and max_chars on /api/extract
x-retain-links, x-retain-imagesDrop or shorten links and imagesIgnored; links and structure are kept
x-proxy-url, x-no-cacheUse your proxy, bypass the cacheIgnored; there is no cache, every read is fresh

If a script depends on a row marked ignored, that part needs a different plan. For token limits, the long-page route below is the closest match. For selectors, you would have to trim the result yourself.

What we measured on real pages

On November 12, 2026 we called the production prefix endpoint, https://imread.org/<page address>, from one machine with no headers set, and recorded the status, the size of the answer in characters and the time. The times include the fetch of the page and one network round trip from our machine, so treat them as one sample, not a benchmark.

PageStatusAnswer sizeTime
Wikipedia, Web scraping20029,722 characters5.3 s
Paul Graham, How to Do Great Work20067,107 characters1.8 s
Tesla 2024 annual report (SEC 10-K)200408,450 characters9.8 s
Jina Reader's GitHub page20023,404 characters2.1 s
react.dev home page2004,905 characters1.4 s

Three things stand out. First, ordinary articles return in a couple of seconds. Second, the Tesla filing is the extreme case: about 408,000 characters in a single answer, which is far more than most models take in one request. On the prefix call that answer arrives whole. If you need to feed a filing to an agent in pieces, use /api/extract with format=markdown and offset, or the MCP tool, which cut the article into parts of about 20,000 tokens, each ending with the offset to ask for next. Third, the react.dev home page returned only 4,905 characters. It is a marketing page whose content is mostly interface, and a reader that extracts the main article will give you little there. That is a good reminder that a reader is not a copy of the page.

We did not run the same pages through Jina Reader. We also did not test imread against pages that need JavaScript, because it is built not to read them, and a quick check of your own pages will tell you faster than any table whether they fall in that group.

Where imread is the better fit

No key and nothing to configure

imread is free to call, with no sign-up and no API key: without a key, one limit of 30 pages a minute per caller applies to everyone. An optional imread Pro key (paid) raises that to 600 pages a minute and adds a batch tool, but nothing here needs one. For a quick script, a one-off research run or a demo, not having to create an account and store a secret is the whole benefit.

WeChat, legacy encodings and SEC filings

imread was built around pages that general readers tend to get wrong. It reads WeChat Official Account articles from mp.weixin.qq.com, decodes Chinese and Japanese pages served in older encodings such as GBK and Shift_JIS, and keeps the tables in SEC EDGAR filings. If your sources include any of these, try imread first. We have not compared its output with Jina's on them, so run both on a few of your own pages before you commit.

Long documents that must fit an agent

Agents have tool-output limits. imread's MCP tool always answers in parts of about 20,000 tokens, because some clients turn away tool answers over 25,000 tokens by default. Each part ends with a line such as For the next part, ask again with offset=N until the last part says [End of the article.]. Parts are consecutive slices cut just after a paragraph or line break, so together they are the whole article. Jina offers a different guardrail, x-max-tokens, which truncates a response to a size you choose. Truncating and paging solve different problems: one caps cost, the other keeps every word reachable.

A reader for humans, not only an API

The same extraction opens in a clean reader where you can listen to the article in a natural voice, translate it into ten languages or copy it as Markdown. When a script gives a strange result, you can paste the same address into the reader and see what a person would see.

One predictable behavior

With two request headers that matter and no engine, cache or timing choices, a given address returns the same kind of answer each time. If you value reproducible output in a pipeline, fewer knobs is a feature.

Where Jina Reader is the better choice

We would rather you pick the right tool than a convenient one, so this is the section to read carefully.

You needimreadJina Reader (per its README)
Pages built with JavaScriptNo: it fetches the HTML and reads thatYes: headless Chrome, with x-wait-for-selector and timing options
Web search for a queryNoYes: s.jina.ai fetches the top five results, and site limits it to chosen sites
Screenshots, raw HTML, image captionsNo: html and screenshot are a 400Yes, including captions written by a vision-language model
PDF and Office filesNo: a link to a file is a 415Yes: PDFs by URL, and Word, Excel and PowerPoint files, including by direct upload
Running it on your own serverNo: hosted onlyYes: Apache-2.0, with a Docker image that runs stateless by default
Sites that block scrapersNo: such pages come back as errorsOptions for proxies, cache bypass and a browser engine
Structured chunks for RAGNo: one article, or parts by offsetYes: x-markdown-chunking splits by heading or block
Higher volume30 pages a minute per caller (600 with an optional Pro key); 9 second fetch budget; 16 MB capHigher quotas with a key; see Jina's pricing page

The first row matters most. A growing share of sites render their content in the browser, and a reader that only fetches HTML will return little or an error for them. imread is honest about this: such pages come back as errors or nearly empty, and the sensible fallback is a tool that runs a browser. If most of your sources are single-page apps, start with Jina Reader or a browser-based tool instead of imread.

Other routes besides a hosted reader

Looking for a Jina AI alternative does not have to mean another hosted API.

  • Self-host Jina Reader. Its README documents docker pull ghcr.io/jina-ai/reader:oss, with an HTTP/1.1 port at 8081 and an h2c port at 8080. With no configuration the container is stateless: every request hits the live URL, with no cache and no rate limiting. You keep full control of what is logged and which proxy is used, and you take on running Chrome and LibreOffice yourself.
  • Extract locally. For static pages, a library such as Mozilla's Readability on a parsed DOM gives you article extraction inside your own process. imread is built on the same library, with extra handling for encodings, tables and WeChat pages. The trade is that you own the fetching, the encodings and the edge cases.
  • Drive a browser. A headless browser reads anything a person can, at the cost of memory, time and maintenance. It suits a small number of hard pages better than a high-volume crawl.

A hosted reader like imread is the cheapest way to start, and a self-hosted one is the cheapest way to stay in control. Many projects use the first and keep the second as a plan B.

Questions people ask

Is imread free, and does it need an API key?

It is free, with no key and no sign-up. It is offered as is, without an uptime guarantee, and limited to 30 pages a minute per caller; an optional imread Pro key raises that to 600.

Can imread read pages that need JavaScript?

No. It fetches the page's HTML and extracts the article from that. Pages that build their content in the browser, need a login or block automated requests will fail or come back nearly empty.

Does imread store the pages I read?

No. Each request fetches the page afresh. Each read writes one anonymous log line with the kind of page, the kind of client and how it went, without the address. A read that crashes logs its address so the problem can be fixed.

Is the output identical to Jina Reader's?

The envelope is the same: the text layout and the JSON fields. The content inside is not guaranteed to match, because the two services extract the article with different code. Compare both on your own pages.

Can I use imread from Claude Code, Codex or Cursor?

Yes. The MCP server has one tool, extract_article, which takes a url and optionally format, offset and max_chars. Setup steps for each client are on that page.

How to choose, step by step

  • List your ten most common source pages. Open each in a normal browser with JavaScript off. If the article is still there, imread can read it.
  • If your sources include WeChat, GBK or Shift_JIS pages or SEC filings, run those through both readers and compare the text.
  • If you need search, screenshots, uploaded files or your own server, use Jina Reader for those paths. Keep imread for the article path if the no-key setup suits you.
  • If you route by page type, send known-static article URLs to imread and everything else to Jina Reader, and fall back to the other when one returns an error.
  • Whatever you pick, handle errors. Both services fail on some pages, and a pipeline that expects that behaves better than one that does not.

How we checked this, and what we did not check

The imread behavior in this guide is taken from its code and from calls to the live service on November 12, 2026: the answer layout, the JSON fields, the 400 for unsupported formats, the 415 for file links and the sizes and times in the table. The Jina Reader description comes from the README in its GitHub repository, github.com/jina-ai/reader, read on the same day. We did not call Jina's hosted API for this article, we did not measure its speed or quotas, and we did not compare extraction quality between the two. Treat the Jina column as a pointer to its documentation and check the current README and pricing page before you rely on a limit.

Try imread on a page you care about: open it in the reader, or read the API docs for the parts, errors and limits. imread is an independent project and is not affiliated with Jina AI.

All posts →