本文来源:https://mp.weixin.qq.com/s/l_tEYZ5f3TduN0sRdp9zIg
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
我先用GitHub的API把jonashonecker/feedpaper这个仓库的提交记录拉了一遍。33次提交,全挤在2026年8月9日和8月10日这两天里,第一天24次,第二天9次,之后就没再动过。Python包底下九个文件,我把每个文件的行数一个个取下来加了一遍,659行。仓库里没有LICENSE文件,接口返回的license字段是null。星标25个,fork零个。
作者自己在博客里写得很直白:借AI的帮忙,几分钟就搭起来了。
这段代码要干的事,小到有点可爱——他不想在手机上读博客了。躺沙发上、坐院子里晒太阳的时候刷工艺博客,眼睛累。这个念头搁了很久没动,直到他看见一台叫X4的小机器。
先说清楚那台机器
Xteink X4,4.3英寸电子墨水屏,没有触摸,没有背光,出厂不带任何书城。官网列的规格是114×69×5.9毫米,650毫安时电池,ESP32芯片,Wi-Fi只有2.4G,随机送一张16GB的microSD卡,能认的文档格式只有EPUB和TXT两种,规格表最后一栏明明白白写着“不兼容第三方应用”。
售价是69美元。
我今天开这个机型页的时候,顶上挂着一条公告:X4已售罄且停产,官方建议去看X3或者X4 Pro。X3同样69美元,3.7英寸;X4 Pro标价99美元。所以这篇里浮现的X4售价,是它退场时的定价,不是现在能买到的价。
顺手记一笔核到的毛边:同一个机型页上,宣传图那块写的是“Featherlight 74G”和“219 PPi”,往下滚到Specifications规格表,变成了77g和220 PPI,底下FAQ又回到74g。差3克、差1个PPI,不影响谁买不买,但一家把“极致便携”当卖点的厂商,自己页面上的重量数都对不齐,这个细节挺说明难题。
更要命的是,那张规格表里“Resolution”这一栏填的是“220 PPI”——PPI是像素密度,不是分辨率。整张官方规格表从头到尾没给过显示屏的像素数。480×800这个数字,我是在CrossPoint固件的使用者手册里翻到的,那篇教人做休眠壁纸的段落写着:X4用480×800,X3用528×792。拿480×800去算对角线像素数除以4.3英寸,得出来大约217,跟官网标的220差着一点。至于这块屏用的是E Ink哪一代面板,Xteink没写,我也没查到。E Ink官网的Carta 1100/1200品牌页上,4.3英寸确实在它列出的尺寸里,但那一页只给相对数据——比较度比Carta 1000提高15%,响应速度快20%——没有任何一个绝对值。
还有一件事得说明:Xteink的服务条款里写“适用香港法律”,但官网上找不到注册公司名和注册地址,关于笔者、联系笔者、隐私政策三个页面我都翻了,没有。
刷进去的那套固件比机器本身有名
这台机器真正的看点不在硬件参数,在它被人刷成了什么。
CrossPoint,开源的电子书电纸书固件,MIT许可证。仓库2025年12月3日建的,到今天7070个星标、1441个fork、207位贡献者、1146次提交,还挂着449个未关闭的issue。最新的v1.5.0是8月7日发的,我拉发布接口数了一下附件下载量,26718次;上一个稳定版1.4.1是87460次。这个量级,对一台停产的69美元小机器来说不算小。
它给X4补的东西,基本上是原厂固件的空白:EPUB 2/3渲染、断字、字距、脚注、书签、StarDict词典查词、KOReader进度同步、网页版文件传输、WebDAV、OPDS书库浏览、从GitHub直接OTA升级、24种界面语言。原博客作者试了几秒原厂固件就受不了了,三次点击刷完CrossPoint,拷进一本无DRM的电子书,能读。
这个仓库还有个细节值得一提:根目录躺着CLAUDE.md和AGENTS.md,.claude/skills底下放了五份写给AI的编码规范,堆分配纪律、控制流清晰度、硬件抽象层边界各一份。一款跑在ESP32-C3上、要跟内存斤斤计较的嵌入式固件,社区已经把“怎么让AI改这份代码”写成了成文规矩。
他主动扔掉了三类订阅源
回到feedpaper这条链路。流程本身没什么玄机:调Feedbin的API取未读条目,标记已读,攒成一个EPUB,拷进设备。仓库里的feedbin.py写死了每次最多取100个条目、每次最多标记1000条已读,一共92行。
真正有意思的是他往外扔的东西。
仓库里有一份文档就叫《为什么排除某些博客有协助》。作者点名三类源不该进这份报纸:链接摘要类的newsletter,一整篇全是URL,电子墨水上点不动链接,那一章就是一串死名单;视频和播客源,正文几乎没有可读文字,内容本体是个播放器;重度依赖JavaScript渲染的博客,工具为了离线干净会剥掉脚本,出来的章节是空的。被排除的源不会被标已读,还留在Feedbin里,作者的原话是“你可以在合适的地方读它们,通常是手机”。
这句话才是整件事的分寸所在。他没有宣称把手机上的信息流搬空了,他只是把其中一类内容——长文——挑出来搬到了一块不发光的屏上,剩下的照旧在手机上。
还有个更细的坎:很多RSS源只给摘要,第一段加一个“阅读全文”。feedpaper的处理是先用内联内容,判断“看着像被截断”就转去调Feedbin自带的全文提取接口,那个接口背后是Mercury Parser,Feedbin在服务端跑,工具自己一行网页都不抓。判断标准写得很土——正文短于大约500个字符、含有“read more”或者德语的“weiterlesen”这类字样、以省略号结尾,任意一条命中就走提取。文档自己承认这套启发式会误判,短文章可能被白白多请求一次。
同样的事Calibre干了十八年
读到这儿我第一个反应是:这轮子不是早有人造过吗。
Calibre有个叫“抓取新闻”的能力,靠的是一套叫recipe的Python脚本机制。我去它的GitHub仓库拉了master分支的树,recipes目录底下1097个.recipe文件,一个文件对应一个新闻源,从大报到小博客都有人写好交上去了。抓完直接转成电子书,官方文档里那句话写得很随意——“如果你的电纸书连着,它会被直接放进电纸书,而不是进书库”。这套代码在Git历史里最早能追到2008年4月。当前版本9.13.0,文档页脚表明的最后更新日期是2026年8月7日。
KOReader那边也有现成的news downloader插件,喂它一个RSS或Atom地址,它给每个源单独建一个文件夹,能规格每次抓几条、要不要抓全文。这插件的wiki页面2019年就写好了。
那feedpaper图什么。
差别实际上很具体。Calibre的recipe是站点导向的——你要读哪家媒体,就找哪家的脚本;feedpaper是账号导向的,它接的是“我在Feedbin里没读完的那一堆”,未读状态、全文提取、稍后读的心智,全在Feedbin那边现成放着。KOReader的插件跑在电纸书上,可X4上跑的是CrossPoint不是KOReader,用不了。至于Calibre,装一套完整桌面软件、维护recipe、再对上设备,对一个只想每天早上拿到一份未读长文合集的人来说,太重了。
往难听了说,feedpaper能存在,很大程度上是因为写它只要两天。十八年前造这个轮子是件正经工程,现在它掉到了“周末顺手写完”的成本区间里。当造轮子比找轮子便宜,重造就成了默认选项。
这一年到底要花多少钱
Feedbin的定价页我自己开了一遍:一个套餐,全能力,月付7美元,年付70美元,页面上标着“省14美元”,30天免费试用不用绑卡。它不是免费RSS服务,这是这条路的入场券。
加上一台69美元的X4,第一年139美元。按欧洲央行8月14日的参考汇率1美元兑6.7413人民币算(我从Frankfurter接口取的),937块。第二年往后只剩Feedbin的70美元,472块,一年。
937块摊到365天,两块五毛七一天。
这里得把话说清楚:我手上没有这台机器,上面所有关于它的话都来自厂商规格表、固件仓库和作者那篇博客,不是我摸出来的。软件那头我能核的都核了——提交数、行数、许可证、recipe个数、定价——硬件那头,480×800是从固件文档里倒推的,电池标称650毫安时对应“每天读一到三小时能撑十四天”这个说法,也只有厂商单维度的口径。
但这笔账值不值,跟机器好不好实际上关系不大。937块买的不是一块屏,是把手机留在家里出门那一次的可能性。它只对一种人成立:本来就在付费用RSS,每天有二三十篇长文积在未读里,而且心里很清楚自己每天在手机上浪费掉的是哪半个小时。这种人不用别人替他算账,他多半自己也能写出那659行代码。剩下的人买回去,两周之后它就是抽屉里一块4.3英寸的砖。
Source: https://mp.weixin.qq.com/s/l_tEYZ5f3TduN0sRdp9zIg If there are any copyright issues, please inform us at 13951431913#139.com or submit a comment on the right.
I first pulled the commit history for the jonashonecker/feedpaper repo from GitHub’s API. Thirty-three commits in total, all squeezed into two days — August 9 and 10, 2026 — with 24 on the first day and 9 on the second. No activity since then. The Python package contains nine files, and adding up the line counts from each yields 659 lines. There’s no LICENSE file in the repo, and the API response’s license field is null. Twenty-five stars, zero forks.
The author explicitly states in his blog post: “With the help of AI, I got it up and running in a few minutes.”
The purpose of this code is charmingly specific: he no longer wanted to read blogs on his phone. While lying on the sofa or sitting in the yard soaking up the sun, reading craft blogs was straining his eyes. This idea had been on the back burner for a while until he saw a device called the X4.
先说清楚那台机器
The Xteink X4 features a 4.3-inch e-ink display with no touch functionality and no backlight. It arrives without any pre-installed bookstore. The official specs list the dimensions as 114×69×5.9 mm, battery capacity as 650 mAh, ESP32 chip, 2.4G Wi-Fi only, and a bundled 16GB microSD card. It supports only EPUB and TXT formats, and the spec sheet explicitly states “Not compatible with third-party apps.”
售价是69美元。
When I visited the model page today, there was an announcement at the top: The X4 is sold out and discontinued; consider checking out the X3 or X4 Pro instead. The X3 is similarly priced at $69 with a 3.7-inch screen, while the X4 Pro costs $99. Therefore, the price mentioned for the X4 in this article is its discontinued price, not what’s currently available for purchase.
A quick note on inconsistencies noticed: On the same product page, the promotional text states “Featherlight 74G” and “219 PPi”, but the Specifications table lists 77g and 220 PPI, and the FAQ reverts to 74g. While a 3g difference and 1 PPI won’t affect a buying decision, such inconsistencies in a company touting “ultimate portability” are telling.
More critically, the “Resolution” field in the spec table is labeled “220 PPI” — which is pixel density, not resolution. The entire official page fails to provide the actual pixel count of the display. The 480×800 value I found in the CrossPoint firmware user guide (in the section about sleep walls: X4 uses 480×800, X3 uses 528×792). Calculating the diagonal pixels from 480×800 divided by 4.3 inches approximately equals 217, close to but distinct from the advertised 220. Regarding the specific E Ink panel generation used, Xteink doesn’t specify, and I couldn’t find confirmation. E Ink’s Carta 1100/1200 branding page does list 4.3 inches, but only provides relative data (e.g., “15% better contrast than Carta 1000”) — no absolute values are given.
Additionally, Xteink’s terms of service state “Governed by Hong Kong law”, but the website lacks a registered company name or address; neither the About, Contact, nor Privacy pages contain this information.
The firmware running on this device is more notable than the hardware itself.
The real attraction of this machine isn’t its hardware stats, but what it has been flashed into.
CrossPoint is an open-source e-book e-reader firmware under the MIT license. The repository was created on December 3, 2025, and as of now has 7070 stars, 1441 forks, 207 contributors, and 1146 commits. There are 449 open issues. The latest release, v1.5.0, was issued on August 7, with 26718 downloads recorded for the attachment. The previous stable version, 1.4.1, had 87460 downloads. For a discontinued $69 device, this is a substantial user base.
CrossPoint adds functionality that the original firmware lacks: EPUB 2/3 rendering, hyphenation, inter-letter spacing, footnotes, bookmarks, StarDict dictionary lookups, KOReader progress synchronization, web-based file transfer, WebDAV, OPDS catalog browsing, direct OTA upgrades from GitHub, and support for 24 languages. The original blogger reports that after suffering through the original firmware for a few seconds, he completed the three-click installation of CrossPoint and loaded a DRM-free book, which read perfectly.
An interesting detail about this repository is the presence of CLAUDE.md and AGENTS.md in the root directory, along with five coding style guides for AI under .claude/skills (covering topics like heap allocation discipline, control flow clarity, and hardware abstraction layer boundaries). For a firmware running on an ESP32-C3 that must strictly manage memory, the community has already codified guidelines on how AI should approach modifications.
The author intentionally excludes three types of subscription sources.
Returning to the feedpaper pipeline, the process is straightforward: fetch unread items via the Feedbin API, mark them as read, compile them into an EPUB, and transfer the file to the device. The feedbin.py script hardcodes a limit of 100 items per request and a maximum of 1000 items to mark as read, totaling 92 lines of code.
What’s truly fascinating is what the author omits.
The documentation explicitly states that certain sources should not be included in this paper: link-only newsletters (which appear as a list of URLs that cannot be clicked on an e-ink display), video and podcast sources (where the main content is audio/video, not text), and blogs that rely heavily on JavaScript rendering (which results in blank chapters after stripping the scripts for offline reading). These excluded sources are not marked as read and remain in Feedbin, allowing the user to consume them elsewhere (typically on a phone).
This decision encapsulates the entire philosophy of the project. The goal isn’t to replace mobile browsing; rather, it’s to extract a specific subset of content — long articles — and present them on a non-emissive screen, while maintaining the status quo for other media types.
There’s also a more technical nuance: many RSS feeds only provide summaries (often truncating after the first paragraph with a “read more” link). feedpaper attempts to detect this by checking if the main content is shorter than ~500 characters, contains phrases like “read more” or “weiterlesen”, or ends with an ellipsis. If detected, the full article is retrieved via Feedbin’s built-in extraction API, which utilizes Mercury Parser. The author acknowledges that this heuristic is imperfect and may result in unnecessary full-text requests for short articles.
This functionality isn’t unique; both Calibre and KOReader offer similar capabilities.
Calibre has a “fetch news” feature that relies on a system of Python scripts called “recipes”. The GitHub repository for Calibre contains 1097 .recipe files, each corresponding to a specific news source, ranging from major newspapers to individual blogs. These scripts allow for the direct conversion of web content to e-book format, with the official documentation noting that if an e-reader is connected, the file can be transferred directly without saving to the library. This feature dates back to April 2008; the current version is 9.13.0, last updated on August 7, 2026.
Calibre features a "Fetch News" capability driven by a system of Python scripts known as recipes. I cloned the master branch of its GitHub repository, which contains 1097 .recipe files corresponding to various news sources, ranging from major newspapers to niche blogs. This mechanism allows users to grab news content directly into electronic book format. The official documentation notes, "If your e-book reader is connected, the book will be sent there rather than added to the library." The code for this feature is traceable back to the initial commit on April 24, 2008. The current version is 9.13.0, with the last update to the documentation recorded on August 7, 2026.
KOReader also offers a news downloader plugin that functions similarly, downloading content from RSS or Atom feeds into separate folders.
那feedpaper图什么。
The critical distinction lies in the target audience and simplicity. Calibre recipes are site-specific — you need to configure a script for each media outlet you wish to follow. feedpaper, on the other hand, is account-specific; it simply retrieves the unread articles from your Feedbin account, leveraging the existing unread status and full-text extraction capabilities of the service. Furthermore, while KOReader’s plugin is powerful, it requires running KOReader on the device, which is not natively supported on the X4 (only CrossPoint is). For someone who simply wants a daily digest of long-form articles without setting up complex site-by-site rules, feedpaper provides a much simpler solution.
In essence, the existence of feedpaper is justified by the fact that creating a custom solution is now faster than finding pre-existing ones. When writing 659 lines of code is quicker than searching for and configuring existing tools, duplication becomes the more efficient choice.
The financial breakdown for the first year is straightforward: a $7 monthly subscription to Feedbin (or $70 annually with a $14 discount), plus the $69 cost of the X4 device, totals $139 for the first year. Subsequent years would cost only the $70 Feedbin subscription. Converting this to RMB at the ECB’s rate of 1 USD = 6.7413 RMB, the first year costs approximately 937 RMB, which breaks down to about 2.57 RMB per day.
The Feedbin pricing page offers a single-tier plan with full access to its features for $7 per month or $70 per year (saving $14 annually). It provides a 30-day free trial that does not require a credit card. This service acts as an entry point into the subscription-based RSS management landscape.
It is important to note that I have not personally tested the X4 device; all hardware information is derived from the manufacturer’s specifications, firmware repositories, and the original blog post. While I have cross-verified the software data (commit counts, line numbers, licenses, etc.), the hardware reliability and battery life claims are purely speculative based on the advertised numbers (650 mAh battery lasting 14 days for 1–3 hours of reading per day).
Breaking down the cost further, a yearly subscription of $70 spread across 365 days amounts to approximately $0.19 per day.
It is important to clarify that I do not own the device referenced. All information regarding it is derived from the manufacturer's specifications, firmware repository, and author's blog post. While I have independently verified software metrics such as commit counts and recipe numbers, the hardware data (like the screen resolution of 480×800 and battery capacity) is extrapolated from the firmware documentation rather than physical measurement.
Ultimately, the value proposition boils down to this: the $139 investment buys you the opportunity to leave your smartphone at home while still enjoying long-form content. This is a viable option for specific readers — those who already pay for RSS, accumulate dozens of unread articles daily, and are conscious of the time wasted scrolling on mobile. For these individuals, the price per day is negligible compared to the benefit of focused reading. For everyone else, the X4 may quickly become a paperweight.
#Xteink #电子墨水 #RSS #Feedbin #CrossPoint #EPUB
相关阅读



