Stringers
Edwardxlai/easyread · Edwardxlai

A "fed up" rant post, 9 days, 860 stars: EasyRead uses your Claude quota to translate papers

What it is A free, open-source desktop paper reader that translates English PDFs page by page into typeset Chinese, with the original page shown alongside, lets you highlight, take notes, and ask AI, and runs translation directly on your existing Claude Code or Codex subscription quota.

Why it took off It has a difference you can state in one line: it uses the subscription you already pay for, with no API key to configure, plus a reading experience that discards the original PDF and reflows it into a book. The author's first-person post naming and roasting mainstream tools in its National Day launch got 107k views, and programmer Yuji and repost accounts brought in two more waves.

BHow you can ride this Stop building yet another paper translator. Take "subscription CLI as backend + reflowed, readable layout + side-by-side original page" and apply it to long documents that are just as hard to read: technical books, EPUBs, patents, annual reports. Launch it on Chinese X with the same first-person rant post.

Grade B: the approach is clear and transferable to other document types; but the paper-translation space is already crowded, "use your subscription quota" is becoming standard for new tools, and its first wave depended heavily on the author being good at writing posts, which you can't easily copy. Circulating in small circles
Data (updated daily) · Repo created 2026-09-30 · 1,364 stars · 7 days +202
GitHub · stars
05001k1.5k10-0110-1810-2911-0911-201.4k
Data 2026-10-01 → 2026-11-20 (51 days, 43 observations). Points before 10-28 come from deeply / archive backfill, not every day has one

The paper-translation space was already crowded; PDFMathTranslate alone has 38k stars. On the first day of National Day holiday, a student in Beijing posted a thread that called out Immersive Translate, DeepL, and Zotero by name. It got 9 stars in 862 days, and seven weeks later reached 1,333. Its features aren't more complete than others'. What it changed is two things: translation runs on the Claude Code / Codex subscription you've already paid for, and what you read isn't an edited PDF but a Chinese book that has been re-typeset. Below: how it took off, where the stars came from, and where to start if you want the same.

A post at noon on the first day of National Day

On 2026 Oct 1 12:14, Edward (@edwardxlaime) on X posted this:

Because my need to read papers has surged lately, and I'm fed up with the bad layout and bad translation of software like Immersive Translate and DeepL, and with the junk literature management of old tools like Zotero, I hacked together a toy in my spare time... Old-school readers like to read the full text in its original form instead of letting a large model just summarize it for them... Can converse right on the reading page (doesn't this crush alphaxiv?). The only downside is it burns your own quota, but not much... Give it a star to show support.

He attached three screenshots. The first is the library: on the left a column of "reading / unread / read," and in the middle three papers with titles translated into Chinese, such as "Direct Preference Optimization: Your Language Model Is Secretly a Reward Model," with a small tag in the top right reading "Claude Code."

By mid-November the post had 107k views, 1,247 likes, and 201 reposts. The comments were mostly the same kind of people nodding along: "If you don't like it, build your own, that's the ultimate romance of tech people," "True, using Zotero's translation is painful for me." Some made requests: "What about OpenCode... and there are free models too. Stop overthinking it, you're just reading papers, not implementing them, no need for Opus!" Others wanted an iPad version, WebDAV sync, and multiple languages. One alumnus of Beijing University of Posts and Telecommunications reposted it saying: "Didn't expect to see a student from BUPT build something this great."

The repo was created the day before the post (9-30). The author's GitHub has only this one repo and 23 followers. His bio lists Beijing Jiaotong University and Beijing University of Posts and Telecommunications. On X he is currently doing enterprise deployment (FDE) at an AI company and has said things like "after I stop doing content creation...", suggesting he may have done content work before. The way that post is written shows someone who understands the content.

What it actually is

A paper reader that runs on your own computer, with installers for Windows, Apple-silicon Macs, and Linux. MIT-licensed and free.

Drag a PDF in, or enter an arXiv ID, DOI, or paper title, and it translates page by page in the background, so you can start reading once the first paragraph is done. The translation isn't pasted back onto the original PDF; it's re-typeset as a Chinese book: body text in Song typeface, formulas re-rendered with KaTeX, and tables turned into the three-line style common in academic papers. On the right you can open the original page, which turns along with where you are reading and highlights the matching passage.

While reading, you can highlight in four colors, take notes, or select several passages and ask AI about them together; AI explanations and any typos in the original found during translation appear in the margin. It has a built-in library, can copy citations in GB/T 7714, APA, and BibTeX formats, and can ask questions across the whole library.

You choose the translation model. Top of the list is Claude Code, followed by command-line tools such as Codex CLI and Cursor CLI, all using your existing subscription quota with no API key needed. It can also connect to DeepSeek, Zhipu (which has free models), Kimi, OpenAI, OpenRouter, or local Ollama. With Claude, the interface shows how much of the 5-hour window quota you've used; when it runs out, it pauses and resumes once the quota recovers (project homepage).

The space was already crowded, and it picked two different angles

ProjectLaunchedScaleHow it differs from EasyRead
PDFMathTranslate2024-0938.5k starsReplaces text within the original PDF layout to produce a bilingual PDF; requires an API; no reader and no AI Q&A
Immersive Translate2022-12Commercial product, 19.8k GitHub starsCan translate anything, PDF is just a side feature; good models require a paid membership
BabelDOC (Immersive Translate team)2024-119,887 starTranslation engine that preserves the original layout
Zotero translation plugins (zotero-pdf-translate and others)From ¥20227.5k to 12.3k starsSelect-to-translate inside Zotero, no re-typesetting
AI4Paper2026-025,270 stars, +715 in the last 7 daysA Zotero plugin plus a web version, a full suite of features, with paid tiers; growing faster than EasyRead
alphaXiv2024Web productAsk AI on arXiv pages, in English, no translation
Drop the PDF straight into Kimi / ChatGPT / Claude—the most usersSummary and Q&A only, no paragraph-by-paragraph translation or original-page comparison

(GitHub,11-18)

In this table, EasyRead is neither the first nor the largest. What made it take off is a difference you can state in two sentences.

First sentence: "Use the quota you already pay for, no API key needed." By 2026, many students and programmers in the Chinese-speaking world had bought Claude or ChatGPT memberships, and the quota often went unused. Tools like PDFMathTranslate require you to apply for and configure a key, and Immersive Translate requires buying another membership. Everyone who repeated it afterward repeated this line: programmer Yuji's 10-03 post said "directly spend the quota of your Claude Code or Codex membership, no API key needed," and another account wrote "the best part is connecting to the Claude Code / Codex CLI subscription, using the quota you already have."

Second sentence: "Throw away the PDF, typeset a book." One English review was titled "The Paper Translator That Throws the PDF Away." Others were trying to fit translations better back into the original PDF layout; it simply dropped the original layout, reflowed the text, and used the original page on the right to solve "I want to check the source." The screenshots look striking as a result, and on X you can tell at a glance that it's a different thing from a bilingual PDF.

But those two points alone wouldn't have carried it. The launch did the rest: the post was first-person, named the tools everyone has used and resents, went out on the first day of the holiday, and ended with a direct ask for stars. The product also held up. The author shipped 7 versions the day of the post, 11-15 versions total by 22, with installers for three platforms and a no-install online preview.

Where the stars came from

DatestarWhat happened
09-300Repo created
10-01291that noon, when the author posted (number from the web archive)
10-03627Yupi (程序员鱼皮) reposted it
10-09862first wave ends
10-09 to 11-16862 → 1,1802 to 21 a day, averaging about 8
11-181,33311-16 repost account "知识分享官" posted it, +153 over two days

(GitHub daily stars; the 10-03 figure comes from a number written in an X post that day)

The first wave was the author's own post, which brought 9 days' worth of 862 stars, about 65% of the total so far. Next came large Chinese AI and programmer accounts: Yupi (全网粉丝 2M+) got 10-03 views and 9,738 likes on 100 post, while several smaller accounts each drew one or two thousand views.

The next five weeks ran on the long tail, seven or eight stars a day. The sources were scattered: a dozen or so AI directory sites and GitHub trending repost bots indexed it automatically, CSDN and GitCode carried machine-generated write-ups, and people shared it on YouTube Shorts, Threads and LinkedIn. The repo also ships a skill for Claude Code / Codex (it lets the AI read your notes in conversation and write answers back beside the passage), so several "Claude Code skill marketplaces" picked it up too.

The second wave came at noon on 11-16, when the X account "知识分享官", which reposts GitHub finds, posted an introduction. The copy was mostly lifted from the README, with "学霸必备" (a must-have for top students) added at the end. It got 23k views and 327 likes. That account's other GitHub posts from the same period usually drew a few thousand views, so this one was its hit.

I found no trace of it on HN, Reddit or V2EX. I could not check whether anyone promoted it on Xiaohongshu, Bilibili or WeChat Official Accounts.

Stars aren't the only signal. People aren't just starring and leaving. In seven weeks the repo picked up 115 forks and 24 contributors, with 170+ issues and PRs combined, and 11-17 had 6 in a single day. Someone added web article import, someone raised the local service's connection queue from 5 to 64, and someone wired in a new model provider. Total downloads across all versions are about 10.7k (GitHub). That figure includes duplicates from multiple platforms and upgrades.

Where it falls short

  • It only reads PDF papers. People are asking for EPUB (issue #139), and replies ask for an iPad version and sync. Books, long reports and standards documents, which are just as painful to read, aren't covered at all.
  • Layout and rendering still have problems. Pseudocode doesn't render and images go missing (#159), dollar signs get read as formula delimiters (#136), cross-page paragraphs occasionally duplicate or break (#140), and model output gets truncated (#168).
  • The Mac installer isn't signed, and someone couldn't open it after installing (#141).
  • People who can't leave Zotero won't move. It draws the crowd that's fed up with Zotero, while AI4Paper, a fuller-featured AI tool in the Zotero ecosystem, gained 715 stars in a week, several times faster.
  • The ceiling is visible. The leaders in this lane are either 38k-star older projects or funded, paid products. A one-person reader reaching a thousand or two stars has already taken the biggest bite out of the "use your subscription quota" pitch.

If it were you

Verdict: B. Worth a try. What it proves is a playbook that can be moved: in a lane that's been picked over, find an angle you can state in one sentence (use your subscription quota, no API key), add a different experience (drop the original layout, reflow the text), and back it with a post written to provoke, and one person with one repo can get close to a thousand stars in a week. The weak points are also clear: the paper-translation lane is already crowded; "subscription quota as backend" is quickly becoming standard for new tools and will stop being a selling point within three to five months (inference); and much of its first wave depended on the author being able to write a good post, which is hard to replicate.

Who it suits: developers who are heavy users themselves and read some painful class of long documents every day; people who can write an emotional, screenshot-heavy Chinese X post or who know a few accounts willing to repost; and people who can ship a release every day and answer issues every day for the two weeks after launch.

Who it doesn't suit: anyone who only wants to build another paper translator and would be competing head-on on features with EasyRead (its community is already filling in features for it, so you'd be fighting PDFMathTranslate all over again); and anyone with no posting channel who is counting on GitHub's organic traffic (EasyRead's natural growth over the last seven weeks has been only about 8 stars a day).

How to do it, ranked by how worthwhile:

  1. Carrying this playbook over to other long documents: English technical books and EPUBs, patents, legal filings, public company annual reports, and standards and specifications. The conditions match the paper case: the readers' English isn't strong enough, existing tools require a key or a separate payment, and the layouts are ugly. Copy the method directly: use a subscribed CLI as the translation backend, reflow the output into a layout suited to that document type (tables for annual reports, claims for patents, chapter structure for books), and keep the original page alongside.
  2. Fill in its weak spots: build a phone or iPad reader that can read the offline web pages EasyRead exports or sync its library. The author doesn't have time for this alone in the short term, and commenters are indeed asking for it. You could also send EasyRead a large PR directly (EPUB support, for example). Its contributor list is public, which makes it the fastest way to build a reputation.
  3. Change the audience: Japanese and Korean readers have the same pain reading English papers, and EasyRead's layout and citation formats were built around Chinese academic habits. But demand and distribution channels in those communities haven't been tested (inference), so this ranks last.

How to get users: write the way its launch post did. Use first person, open with "fed up with XX", name one or two tools everyone uses and resents, say in a sentence how you differ from them, include three screenshots so the "Claude Code" label is visible in the interface, and end with a direct ask for stars. Post on the first day of a holiday or just before the winter break. After launch, reply to every request in the comments and ship a release that same day. Then get the repo link to the GitHub repost accounts and programmer influencers on Chinese X, and include a skill in the repo so the skill marketplaces will index it.

Deadline: launch before mid-January of 2027, in time for winter break. The longer you wait, the staler "use your subscription quota" becomes.

What would mean I read it wrong: (1) EasyRead hit GitHub trending two months in, gaining a few hundred stars a day, and daily growth hasn't dropped back to single digits. That means the ceiling is higher than estimated, and the paper angle still has room. (2) PDFMathTranslate and AI4Paper, both big projects, plugged in subscription CLIs as backends within a month or two. Then that line is no longer a differentiator, and I'd need a different angle for other document types. (3) If Anthropic or OpenAI tighten how subscriptions can be used in third-party tools, the foundation of the whole playbook is gone. (4) Conversely, if someone applies the same playbook to books or EPUB and a single post gets 500+ stars within two weeks, that confirms path 1 is right, and it's time to build it fast.

Data and sources

GitHub: the Edwardxlai/easyread repo page, README, release and download counts, contributors and issue/PR lists (as of 11-17); daily stars (day by day from 10-09, with the 10-01 figure from the web archive and the 10-03 and 10-07 figures from numbers written in X posts); star counts for comparable repos (11-18). X: @edwardxlaime's 10-01 launch post, its screenshots and comments; @yupi996's 10-03 post; @knowledgefxg's 11-16 post; reposts by @iLegend_AI, @YuewenFang and others. Project homepage edwardxlai.github.io/easyread. Review article: heatdrop, "The Paper Translator That Throws the PDF Away". Xiaohongshu, Bilibili, WeChat Official Accounts and linux.do were not checked or could not be checked. Places marked "inference" or "unverified" are inference or unverified.

GitHub repo ↗

Research notes: competitor table, acquisition evidence, full research report
Competitors
NameWhereLaunchedScaleHow it differs from this one
PDFMathTranslate (pdf2zh)Open source2024-0938,495 starGenerates a bilingual PDF by replacing text on the original PDF layout; requires an API; no reader, no AI Q&A, no library, and cannot use subscription quota
BabelDOC (Immersive Translate team)Open source2024-119,887 starA layout-preserving document translation engine, not a reading workspace
Immersive TranslateChrome2022-1219,825 GitHub stars, millions of users (unverified)General bilingual translation, with PDF as a side feature; advanced models are paid; the author's launch post specifically criticized its layout
zotero-pdf-translate / zotero-pdf2zh / Zotero GPTOpen source2022-03 / 2025-01 / 2023-0312,309 / 8,390 / 7,532 starIt plugs into Zotero for inline translation or calls pdf2zh, rather than reflowing the paper into a Chinese book, and it does not use a subscription quota.
AI4Paper (wdcpclover/ai4paper)Open source2026-025,270 stars, +715 in the last 7 daysZotero plugin plus web version, an all-in-one bundle (journal rankings, 240M+ literature search), commercialized; EasyRead is a standalone desktop reader focused on subscription quota and reading experience
gpt_academic / ChatPaperOpen source2023-0371,679 / 20,017 starAn older-generation GPT academic toolkit, mostly summarization and polishing; not a paragraph-by-paragraph reader
alphaXivWeb2024Not verifiedOnline AI Q&A and commenting on arXiv pages; English only, no translation; the author calls it "crushing alphaxiv"
Doc2X / Zhiyun / Scholaread / EasyReader易读 / Super Literature ReaderWebvariousNot verifiedClosed source, charged per page or by subscription; EasyReader易读 shares the name with EasyRead
Upload the PDF directly to Kimi / ChatGPT / ClaudeBig-company features—HugeOnly summarization and Q&A; no faithful paragraph-level translation and no side-by-side with the original page
Agentero (poco-ai) / PaperRead / 读 YiduOpen source2026-07 / 2026-10 / 2026-102,623 / 368 / 3 starAgent-first literature management, reading workspaces, sentence-level alignment; none combine "subscription quota as backend" with Chinese book-style layout
Where users come from
  • Verified Author's own X launch post: 2026-10-01 12:14 @edwardxlaime, 107,161 views, 1,247 likes, 201 reposts; it named 沉浸式翻译 / DeepL / Zotero as the tools to beat and ended with a request for stars; it went out during the National Day holiday; stars went from 0 to 862 within 9 days
  • Verified Reposted by Chinese programmer / AI accounts: @yupi996 (程序员鱼皮) posted 10-03, with 9,738 views and 100 likes; the pitch was "use your Claude Code or Codex membership quota directly, no API key needed"; also @WEEXAILabs 1.2K and @weiwei2018831 with 829 views
  • Verified X repost account "知识分享官": 2026-11-16 @knowledgefxg, 23,323 views, 327 likes, 60 reposts; 11-16 to 11-18, +153 stars over two days
  • Verified Auto-aggregator sites / SEO / GitHub trending bots: indexed by skillget, ai427, aiboss88, 蝶动 AI导航, 有招, PyShine, heatdrop, hysenlabs, CSDN gitblog, GitCode blogs, YouTube Shorts, Threads and LinkedIn; "GitHub trending" bot posts on X
  • Verified Agent skill marketplaces: TypingMind Skills, SkillsLLM, agentskillshub and aiya.pw indexed skills from the repo, such as paper-reading
  • Inference Xiaohongshu / Bilibili / WeChat Official Accounts / linux.do: not found or not checkable (TikHub balance used up; linux.do behind Cloudflare)

Edwardxlai/easyread (EasyRead): why it took off

Data as of 2026-11-18. Sources: our GitHub database (repo_profile / repo_day; the 10-01 data point comes from wayback), GitHub API (releases, contributors), live X search (x live / x post / x replies), Jina / Exa web search.

Bottom line

一个北邮学生(现在做 FDE)自己用的论文阅读器,2026-10-01 国庆假期第一天他在 X 中文圈发了一条带吐槽的发布帖(10.7 万浏览、1247 赞、201 转),一周内拿到约 860 星。之后两周有程序员鱼皮这类大号转发,一堆自动聚合站跟着收录。11-16 搬运号「知识分享官」又发了一条(2.3 万浏览),带来第二波,两天约 +153 星。 它戳中的点是用你已经付过钱的 Claude Code / Codex 订阅额度来翻译论文,不用 API Key,再加上把论文重排成「排好版的中文书」(宋体、KaTeX 公式、三线表、原页对照)。这两点都是 PDFMathTranslate、沉浸式翻译那条「在原 PDF 上覆盖译文」的路线没有的。

What it is

  • A local desktop app (Electron shell with a Python backend), shipping installers for Windows, Apple Silicon Mac and Linux, MIT-licensed open source. Homepage edwardxlai.github.io/easyread/
  • How it works: drag in a PDF, or enter an arXiv ID, DOI, title, or an OpenReview / ACL / bioRxiv link. It translates page by page in the background, so you can start reading once the first paragraph is done. Body text is set in a Song-style serif, formulas are re-rendered with KaTeX, and tables become three-line tables. On the right, the original page opens and turns pages with your reading position, highlighting the current paragraph.
  • While reading you can underline in four colors, take notes, and select several passages to ask the AI about together. The AI's explanations, and any typos in the original it finds during translation, appear in the margin. It has a built-in library, can copy GB/T 7714 / APA / BibTeX citations, export to Markdown or a single-file offline web page, and can ask questions across the whole library.
  • Translation backends: Claude Code (recommended) / Codex CLI / Grok Build / Antigravity / Cursor CLI, all of which use your subscription quota, no key needed. It can also connect to APIs from DeepSeek, Zhipu (GLM-4.7-Flash is free), SiliconFlow, Kimi, OpenAI, Gemini, OpenRouter and others, or to local Ollama. It shows how much of your Claude subscription's 5-hour window you've used, pauses translation when the quota runs out, and resumes when it resets (official site).
  • It ships with its own CLI and agent skill: Claude Code and Codex can read your notes in the conversation, write answers back next to the paragraphs, or re-translate specific pages. That's why skill marketplaces such as TypingMind, SkillsLLM and agentskillshub list it as a skill.
  • Release cadence: 10-01 shipped v1.0.0, then v1.0 through v1.2.5 the same day. From 11-15 to v1.5.0 there have been 22 releases in total (GitHub API). Cumulative downloads across all release assets are about 10,700, including repeat downloads across platforms and upgrades. The three latest versions, v1.4.5 / v1.4.6 / v1.4.7, each have roughly 1,300 to 1,500 downloads.
  • The author, Edward (GitHub @Edwardxlai, X @edwardxlaime): his bio says he's a student at Beijing Jiaotong University / Beijing University of Posts and Telecommunications. This is his only repo on GitHub, with 23 followers. On X he says he works as an FDE (enterprise AI deployment) at an AI company, and has written "After I stopped doing content creation...", which suggests he has made content before and knows how to write for an audience.

Growth curve

DatestarNotes
09-300Repo created
10-01291Wayback snapshot; the author's X launch post the same day 12:14
10-03627Figure cited in a post by X user @iLegend_AI
10-07821@TheArthiAI "Trending on GitHub" post
10-09862deeply repo starts daily logging
10-09 to 11-16862 → 1180Natural growth of 2 to 21 stars per day, averaging about 8 stars/day
11-181333Two days after the 11-16 evening "knowledge sharing officer" post, +153
  • The first 9 days brought 862 stars, 65% of the total. That's a cold start driven by a launch post. The following five weeks ran on long-tail traffic, about 8 stars a day.
  • gain_7d 188 / prev_7d 52, status = accelerating, but nearly all of that comes from the 11-16 repost.
  • Stars-to-forks ratio is about 8.5%, with 113 forks and 10 external contributors (TWwWT1, Wang-auspicious, hyskr and others have submitted PRs). People in the issues are suggesting features. The suspect list is empty, so this doesn't look like star farming.
  • The repo is only 49 days old, and based on the pattern that trending only admits repos after about 60 days, it isn't eligible for GitHub trending yet. Once it's been up 60 days at the end of November, if it gains enough stars that day it could make trending for the first time, a possible third wave.

Where users come from (growth levers)

  1. The author's own launch post on X (main driver, confirmed). Posted 2026-10-01 12:14 (https://x.com/edwardxlaime/status/2105511571967668267), 107,161 views, 1,247 likes, 201 reposts, 47 replies. The copy is an emotional rant: "Sick of the junk layout and bad translations from immersive translation and DeepL-type tools, sick of the outdated junk in Zotero-style reference managers... Won't this crush alphaxiv? Only downside is it burns your own quota", ending with "Give it a star to support!" It went out on the first day of the National Day holiday, when students and researchers have time to tinker with tools. The replies compare it to immersive translation and Zotero, and some asked for OpenCode / free model support, multilingual support, an iPad version and WebDAV. Beijing University of Posts and Telecommunications alumni also shared it (@YuewenFang: "Didn't expect to see a BUPT student build something this good").
  2. Repost by a major Chinese AI / programming account (confirmed). @yupi996, Chinese programming blogger Yupi (2M+ followers across platforms), posted 10-03, with 9,738 views, 100 likes and 15 reposts. The pitch: "Use your Claude Code or Codex membership quota directly, no API key needed," plus "Karpathy just recommended ASD-STE100 short-sentence answers." Smaller accounts followed, including @weiwei2018831 (829 views) and @WEEXAILabs (1,243 views).
  3. Auto-aggregator sites and GitHub trending reposts (confirmed, small individually but wide reach). Between 10-01 and 10-24, skillget.dev, ai427.com, aiboss88, Diedong AI navigation, Youzhao, PyShine, heatdrop ("The Paper Translator That Throws the PDF Away"), hysenlabs, code-corey, auto-generated CSDN gitblog articles, GitCode blogs, a Turkish Edumints blog, YouTube Shorts, Threads and LinkedIn either listed or rewrote it. On X there are also "GitHub trending" and "headline trending" bot accounts (@ppmmzzcn, @TheArthiAI). These bring long-tail traffic and SEO, not a spike.
  4. Agent skill marketplaces (confirmed). The repo ships a skill (paper-reading and others), and TypingMind Skills, SkillsLLM, agentskillshub and aiya.pw listed it as a "Claude Code skill," picking up traffic from the 2026-year skill boom.
  5. 11-16 repost account "knowledgefxg" on X (@knowledgefxg), wave two, confirmed. x.com/knowledgefxg/status/2122303440198185365, 23,323 views, 327 likes, 60 reposts, with the "must-have for top students" pitch. The timing lines up with the +153 stars from 11-16 to 11-18. This is a routine GitHub gadget-roundup account, and its usual GitHub posts only get 7,946 views, so this one is clearly its hit. 11-05
  6. Not found: HN, Reddit, V2EX (not searchable on sov2ex), linux.do posts (site search is blocked by Cloudflare, unverified), Xiaohongshu / Douyin (TikHub's old account balance ran out, not checked), media coverage, paid promotion.

Competitors

CompetitorsFormLaunchedScaleHow it differs from EasyRead
PDFMathTranslate(pdf2zh)Open source2024-0938,495 stars (EMNLP 2025 Demo)Replaces text on the original PDF layout and outputs a bilingual PDF. Needs an API key or uses Google Translate. No reader, no AI Q&A, no library, and can't use subscription quota.
BabelDOC (funstory-ai, the Immersive Translate team)Open source2024-119,887 starsA document translation engine that also keeps the original layout in PDF output. It's the foundation of Immersive Translate's PDF Pro.
Immersive TranslateChrome extension / web / app, commercial2022-12GitHub 19,825 stars, millions of users (unverified)A general-purpose bilingual translator where PDF is a side feature, with paid advanced models. It's exactly the tool the author named and criticized in his launch post.
zotero-pdf-translate / zotero-pdf2zh / Zotero GPTOpen-source Zotero plugin2022-03 / 2025-01 / 2023-0312,309 / 8,390 / 7,532 starsLives inside Zotero for selection-based translation or calls pdf2zh. Doesn't reflow the text into a "Chinese book" layout.
AI4Paper(wdcpclover/ai4paper)Open-source Zotero plugin plus a web version, with commercialization2026-025,270 stars, +715 in the last 7 days, growing faster than EasyReadAll-in-one feature suite (full-text translation, knowledge-base Q&A, journal rankings, 240M+ literature search), running on the Zotero ecosystem and its website
gpt_academic / ChatPaperOpen source2023-0371,679 / 20,017 starsThe older generation of "GPT academic toolbox," mostly for summarizing and polishing. Not a reader.
alphaXivWeb2024Not verifiedAsk AI and comment on arXiv paper pages. English only, online, no translation. The author calls it "crushing alphaxiv."
Doc2X, Zhiyun, Scholaread, EasyReader, Chaoneng Literature, WisPaperCommercial web / desktopvariousNot verifiedClosed source, charges per page or by subscription. EasyReader also shares a name with EasyRead.
Drop the PDF straight into Kimi / ChatGPT / ClaudeBig-company features—HugeSummarizes and answers questions only. No paragraph-by-paragraph faithful translation and no side-by-side original page. The author says "old-school readers like the original full text."
New open-source projects in the same period: Agentero (poco-ai), PaperRead (YuriyLin), Yidu (pph-hpp), Paper Between (browser extension), texglotOpen source2026-07 to 102,623 / 368 / 3 stars / — / 127 starsAgentero focuses on "agent-first literature management." PaperRead is a similar workbench from 10-09. Yidu is a side-by-side sentence-level viewer, Windows only. None combine "subscription quota as the backend plus Chinese book-style layout."

Takeaway: the paper translation space is crowded. PDFMathTranslate alone has about 30 times the stars. EasyRead wasn't first and isn't the biggest, but it found a new angle in a space that's already well covered.

Differentiation

  1. Using your existing subscription quota as the translation backend (the main selling point and the main thing people share). 2026 Lots of Chinese developers and students bought Claude Max or ChatGPT Plus and never use up the quota. The line "no API key, no extra spending" is repeated by everyone who shares it: the post from indie dev 程序员鱼皮, the post from @iLegend_AI ("the sharpest part is that it plugs into your Claude Code / Codex CLI subscription and uses the quota you already have"), and the first row of the official site's comparison table. PDFMathTranslate and similar tools require an API key, and Immersive Translate requires a paid membership. No competitor has taken this angle.
  2. "A properly typeset Chinese book" instead of "translation overlaid on the original PDF". The title of the heatdrop review is "The Paper Translator That Throws the PDF Away." It discards the PDF and reflows the content into a web reader (Song-style serif, line width, KaTeX, three-line tables), then solves the "check against the original" problem with the original page on the right and paragraph highlighting. Reading it feels nothing like pdf2zh's bilingual PDF, and the screenshots look good, which suits sharing on X.
  3. Separating translation from explanation, asking AI with context, and letting the AI cite your highlights. It's a reader plus a copilot, not just a translator. The benchmark here is alphaXiv, not DeepL.
  4. Agent-friendly: it ships a CLI and a skill, so Claude Code / Codex can read and write your annotations directly. That got it listed in skill marketplaces a second time.

Were those differences the reason it blew up? Yes, but only half of it. The subscription-quota angle gave "why not just use pdf2zh" a one-sentence answer, and that's the line everyone repeated when they shared it. The other half was distribution: the author has self-media experience, posted with "tired of XX's garbage" to call out mainstream tools by name, timed it for the National Day holiday, and asked for stars right in the post. The product itself is polished: on launch day it shipped 7 versions and 22 releases, with installers for all three platforms and a live demo, so people who came stayed, starred it, and some even sent PRs.

How many people have seen it

  • Circulating in small circles: mostly confined to the Chinese X AI / developer community (the author's post had 107k views, Yupi 9.7K, the knowledge officer 23k) and a pile of auto-aggregator sites. No media coverage, not on HN or Reddit, and not yet on GitHub trending (too new). It's free and open source, so there's no revenue to speak of.
  • Copycats: no repo I could find that clearly copies it. A few similar open-source projects appeared around the same time, but nothing shows they were copying it: PaperRead (built 10-09, 368 stars), Yidu (built 10-24, 3 stars), and Zhijian (November, browser extension). On X, someone replied "I drew inspiration from your project" (@LukasReyu), but I didn't find a finished product. Contributor Wang-auspicious built paper-xray (228 stars) on 10-03: a close-reading prompt, a spin-off rather than a competitor.

What others can borrow

  • "Use the subscription you already pay for" is a general angle for open-source tools in 2026: any desktop tool that calls an LLM (document translation, subtitles, books, code review) can position against older rivals that make you configure an API key, just by using Claude Code or Codex CLI as the backend. This transfers to other document types (books, bid documents, patents, minor languages).
  • In a crowded category, change the experience axis completely: for example, "throw out the PDF and reflow it into a book" instead of making a copy of the original PDF viewer that looks a bit better.
  • How the launch post was written: first person, starts with the pain point, names the mainstream tools it criticizes, states the difference in one line, and asks for stars directly. Post it on a holiday, reply to every user request in the thread, and ship fixes fast.
  • Add a skill and a CLI to the repo: this gets it listed in skill marketplaces and the agent ecosystem.

Still unclear

  • Whether it spread on Xiaohongshu, Bilibili, Douyin, Zhihu or WeChat Official Accounts: I didn't check the old TikHub account balance, which ran out; the Bilibili videos Jina found are for a different product. Whether the developer Yuhu (程序员鱼皮) also promoted it on WeChat or Bilibili is unverified.
  • Whether there was a launch post on linux.do: site search was blocked by Cloudflare, so unverified.
  • Daily star distribution from 10-01 to 10-09: the stargazer API wouldn't return data with our token, so I only have four points, 10-01, 10-03, 10-07 and 10-09.
  • Real active user count: release downloads are about 10.7k, which is only an upper-bound reference; the number of users who run it from source is unknown.
  • Whether it can get onto trending for a third wave after 60 days, and whether a competitor like AI4Paper, with more features and faster growth, squeezes it out.

This article was written by AI from public data and the data we collect daily. Numbers are sourced, and inferences are labeled.