Karakeep: A Bookmark and Web Archive That Tags Itself

QiaoLink Rating · Public Report

The QL Score is shown as a range because the current evidence does not allow us to pin it to a single number. The ends of the range are rounded to whole numbers. The rating is set by comparing the midpoint of the range (63.4) against fixed thresholds; the midpoint is calculated from the unrounded raw scores. The midpoint is only used to assign the rating band; it is not a "true, precise score".

Current recommendation

• If you run your own server and know Docker: worth trying. The software is free and complete, and your data stays with you. Give it a server with at least 2GB of memory, set a memory limit for it, and turn off open sign-ups.

• If you don't want to manage a server: if you save a lot and don't want to organize it by hand, try the cloud Pro plan ($4 a month, with a 7-day refund). If you only save the odd link, your browser's bookmarks are enough.

• If privacy matters to you: according to the cloud privacy policy, automatic tagging and summaries are processed by OpenAI. When self-hosting, the default models are also OpenAI's, but you can switch to a local model such as Ollama, or leave AI off.

Key Rating Drivers

• Many users say publicly that it makes saved pages far easier to find than browser bookmarks, and that it is a good replacement for Pocket, which has shut down: pages are archived in full, searchable, and tagged automatically.

• The project is very active: 53 releases since February 2024, and reported problems get responses. But more than 80% of the code comes from one core developer, who also has another job.

• The self-hosted version has a memory leak that has been open since January 2026. AI tags tend to pile up over time; an official fix in August 2026 has not yet been confirmed by users.

• The export function on the website does not include plain-text notes or images you uploaded yourself; moving everything out requires a command-line tool.

• It has not been shown by independent reviews to be better than comparable tools such as Linkwarden or linkding.

What it is

Once you have a lot of browser bookmarks, you can rarely find anything again, and pages may disappear after a few years. Karakeep is built for that problem: it collects the web pages, notes, images and PDFs you want to keep in one library.

• When you save a page: it captures the title, the text and a screenshot, and can archive the whole page so you can still read it if the original goes away.

• Organizing: AI tags items and writes summaries, so you don't have to sort everything yourself.

• Finding: full-text keyword search; since a release in August 2026, you can also search by meaning, without remembering the exact words.

• Where it runs: a web app, iPhone and Android apps, and Chrome, Firefox and Safari extensions.

• For people who code: an API, a command-line tool, and hooks for AI assistants.

It was released as open source under the name "Hoarder" in February 2024 and renamed Karakeep in April 2025. The code is public (AGPL-3.0 license) and has more than 30,000 stars on GitHub.

AI is optional: without AI, it is still a bookmark tool with archiving and full-text search. To use AI when self-hosting, you add an OpenAI API key or connect a local model running on your own machine; in the cloud service, AI is only included in the paid Pro plan.

Two ways to use it, with very different barriers

Prices and limits are from the official pricing page, verified on 2026-10-31.

How well does it work?

Most users are positive. The following are things users have said publicly on Hacker News; QiaoLink has not tried the product itself.

• One user moved about 1,000 Firefox bookmarks into it and said: "I didn't realize just how awful the discoverability of old bookmarks was until I started using karakeep to search for items."

• After Pocket shut down in 2025, one user switched to Karakeep and said "it's been a perfect replacement"; another reported "No issues with the import" of their Pocket saves.

• One user runs tagging and summaries on a local AI model and says that now, "when I need to find things again it's never been easier."

• One user who had tried Linkwarden and linkding wrote that they "then settled on KaraKeep"; another chose it over Linkwarden because it has a native Android app.

• A blogger who has run it on a Raspberry Pi for about a year wrote that the AI tags "are consistently better than the ones I'd have bothered to write myself, mostly because I would not have bothered to write any." A writer at the tech site XDA found its "sorting and context features are vastly superior" to browser bookmarks.

There are also complaints:

• The built-in reader is not as smooth as reading the original page; in XDA's testing, it "often fails to properly display posts from Instagram and Imgur."

• Some sites with anti-bot protection can't be captured. A release in October 2026 was designed to improve this, but the result has not been independently checked. For paywalled articles that need a login, the server can only save the opening excerpt.

• AI tags sometimes go wrong: one user reported that with certain local models, prompt text leaked into the tags, or a single bookmark got more than 90 tags.

• Some say it does not replace a dedicated read-it-later app, and some have moved on to the simpler linkding, the lighter Readeck, or other tools.

AI tags tend to pile up. This is the most common complaint from users:

• One user with about 2,000 bookmarks found the AI had created "about 10k tags," and the tags page froze.

• Another, who imported 16,000 items from Pocket, wrote: "The AI tagging is AMAZING but it generates a LOT of tags and that makes the tag management screens in Karakeep difficult or impossible to use."

• One user who tried limiting tags with a custom prompt said it "doesn't really work."

In August 2026, the developer changed the approach: before tagging a new bookmark, Karakeep now looks up similar saved bookmarks and suggests reusing their tags. The developer says this means "a lot less tag sprawl." The change relies on the same feature that powers search by meaning (on by default with OpenAI; it needs extra setup with local models), and there is no user feedback yet on how well it works. One user said they would more likely just turn automatic tagging off.

How good the AI tags are also depends on the model you use. The official documentation says so too. There is no independent review yet of which models work well.

Is it stable?

The project is actively maintained: 53 releases since February 2024, the latest a security update on October 22, 2026; code changes go through automated checks; and the developer responds in issue reports. In 2026, a new version has come out about every 2–3 months, and the October release opened with "sorry for the long wait."

Development depends mainly on one person. On GitHub, 82.7% of code contributions come from a single developer, and only seven people have contributed 10 times or more. The developer's GitHub profile shows that they also have another job. The project is open source, so others could in principle take over, but there is no established second tier of maintainers today.

But if you host it yourself, there are two known problems to watch (from user reports on GitHub; QiaoLink has not reproduced them):

• Memory leak: users first reported in January 2026 that its memory use keeps growing to several gigabytes. The developer traced it to periodic health checks from proxy or monitoring services, but there is still no official fix. One user's server ran out of memory and went down because of it. Workaround: set a memory limit for the container and turn off health checks against it.

• Large imports: one user imported about 1,500 links at once, and it used up first 2GB and then 4GB of memory. This has not been fixed either.

Upgrades sometimes also require following migration steps, such as replacing the browser component, and an upgrade occasionally introduces new problems. Adding bookmarks through the Android app's share menu is also unreliable, a problem first reported in 2024 that is still open. With the cloud service, server problems are handled by the operator, not by you.

AI and privacy

• Cloud service: according to the privacy policy, automatic tagging and summaries are processed by OpenAI, payments are handled by Stripe, and data is stored mainly in Europe. The policy does not say how long OpenAI keeps the data. The browser extension has an option to check whether the current page is already saved; it is off by default.

• Self-hosted: your data stays on your own server and the developer cannot see it. The default AI models are OpenAI's (active only once you add your own OpenAI key), and you can switch to a local model such as Ollama. If you don't want content to leave your machine, switch both the tagging model and the model used for search by meaning to local ones.

• If you self-host on the public internet:

  • anyone can create an account by default, so you need to turn sign-ups off in the settings;

  • the official documentation warns that untrusted users could use your server to reach other websites or even probe your internal network. So don't open a self-hosted instance to strangers.

• If you want to leave: the export on the website only includes link bookmarks; plain-text notes and images you uploaded yourself can't be exported (reported by a user in October 2026 and still unresolved). To move everything out, you need the official command-line migration tool (uploaded files are limited to images and PDFs), which is not friendly for people unfamiliar with the command line. The cloud service keeps your data for 30 days after you cancel; if you self-host, the data is already yours, and automatic backups are available.

Is it worth paying for?

Prices and features as shown on each product's own website (2026-10-31). Information on the other tools comes from their own descriptions; QiaoLink has not tested each item.

• Compared with browser bookmarks: full-page archiving, full-text search and automatic tags are a real step up, so it is worth it.

• Compared with similar tools: its feature set is very complete (AI tags plus search by meaning, apps on every platform, and an API), but no independent review has shown it to be better. One comparison article of the three tools calls it "the most capable of the three" with the best mobile apps, but also "the heaviest"; linkding is the lightest, and Linkwarden is closest to Raindrop.io. One user who switched to Readeck saw memory use drop from about 700MB to about 50MB. Choose based on your needs.

Who it suits, and who it does not

Suits

• People with a home server or cloud server who know Docker and want to keep their data under their own control: the software is free and complete.

• People who save a lot, don't want to organize by hand, and are willing to spend $4 a month: use the cloud Pro plan and skip server upkeep.

• People who want to plug their saved items into their own automations or AI assistant: the API and tools are all there.

Less suitable for

• People who only save the odd link: browser bookmarks are enough, and the free cloud plan has just 50 bookmarks and no AI.

• People who can't maintain a server and don't want to pay.

• People whose server is short on memory: the lighter linkding or Readeck is a better fit.

• People with many notes and uploaded images who can't use the command line for backups: the website export doesn't include them.

• People who mainly want a comfortable way to read long articles, or who often save paywalled articles: it is more of an archive than a dedicated read-it-later app.

• People with saved content they don't want to send to OpenAI and who don't plan to run a local model.

• People who want to let strangers use their self-hosted instance.

Practical advice if you self-host

• Set a memory limit for the container; if a proxy or monitoring service health-checks it, turn those checks off.

• If it is on the public internet, turn off new sign-ups in the settings.

• Read the release notes before upgrading and follow any migration steps.

• Switch to a local AI model if privacy matters to you.

• Make full backups regularly: notes and uploaded images need the command-line migration tool or automatic backups, because the website export leaves them out.

Why this rating

• The core features are dependable and users like it: saving, archiving and finding things work for many long-term users, and feedback is positive.

• But the most important questions still lack independent evidence: there is no independent review of how accurate AI tagging and search by meaning are; the effect of the official fix for tag pile-up is unknown; and comparisons with Linkwarden and similar tools are scattered personal accounts that point both ways.

• Stability has a clear weak spot: for self-hosters, the memory leak and the large-import problem are both marked high priority by the developer but have gone unfixed for a long time.

• Long-term reliance needs care: development depends mainly on one person, and notes and uploaded images can't be exported from the website, so switching tools is not easy.

• The barrier differs a lot between the two kinds of user: the cloud service works as soon as you sign up, but AI requires paying; self-hosting is free but requires Docker skills and handling upgrades and problems yourself.

• That is why this is "adoption under limited conditions": it is worth using for people who can self-host or are willing to pay for the cloud plan, and who really do have a lot to organize; others may not need it.

• What could lead to an upgrade: if the memory problems are fixed and independent reviews show its quality is better than that of comparable tools, there would be grounds for an upgrade.

Where the evidence comes from

• Directly verified by QiaoLink (E2):

  • the official homepage, pricing page and apps page (2026-10-31);

  • the cloud privacy policy (effective 2026-02-15) and terms of service;

  • the official documentation: installation, configuration, AI costs and security considerations;

  • GitHub repository data: stars, creation date, the release history of 53 versions, contributor statistics, automated checks and issue counts;

  • the code change that raised the free plan's limits (2026-10-11) and an earlier web archive snapshot of the pricing page (2026-07-28);

  • the official documentation on data migration and import;

  • the operating company's record at Companies House in the UK;

  • the websites of Linkwarden, linkding and Raindrop.io.

• Independent external evidence (E3):

  • public feedback from more than twenty users on Hacker News, both positive and negative;

  • personal blogs (a year of use on a Raspberry Pi; a switch to Readeck), an account of use at the tech site XDA, and a reader comment on LWN;

  • a comparison article covering Karakeep, Linkwarden and linkding;

  • problem reports filed by users on GitHub (memory leak, large imports, tag pile-up, missing exports, failed page captures and others).

• Self-reported by the developer (E1):

  • descriptions of new features and their effects in the release notes;

  • the estimate of AI tagging costs;

  • the developer's explanation of the cause of the memory problem.

• Still missing (E0):

  • a quality review of AI tags, summaries and search by meaning, and of the August 2026 tagging change;

  • a systematic independent comparison with similar tools;

  • an uptime record for the cloud service;

  • a timeline for fixing the memory problem;

  • app store ratings and reviews;

  • how long OpenAI keeps data processed for the cloud service;

  • any plan for others to take over from the core developer.

• Verification limits: we did not install, sign up for or try Karakeep for this rating; neither the self-hosted version nor the cloud service was tested. The user experience and AI results described here come from users' public feedback and have not been directly verified by QiaoLink.

QiaoLink did not sign up, pay or contact the developer in the course of this rating. Evidence is current as of 2026-10-31.