One search across open research · no tracking · your own key
Finding a paper is largely a solved problem. Deciding whether it is worth your afternoon is not. This is a reading tool built on top of an open index: it searches, and then it helps you weigh what came back.
It is built and run by one person in Antwerp, Belgium. There is no company behind it, no investors, no advertising and no tracking. It runs on hardware I own.
Deliberately without a total. Every aggregator quotes a headline number for its index, and none of them is a number I have counted myself.
OpenAlex supplies the index and Unpaywall finds a free copy where one exists; those two do nearly all of the work. Records from arXiv, PubMed, Zenodo, bioRxiv and institutional repositories are reached through that index rather than by asking each of them separately. The Smithsonian collections are a much smaller addition, connected by hand, and they are here because research does not stop at journal articles. You can also pull in your own work through your ORCID iD, which sidesteps the indexing lag that affects preprints, Zenodo records and books.
Paste a DOI, an OpenAlex ID, an arXiv number or a PubMed ID and it goes straight to that record. Type a person's name and you get their papers, not the papers that happen to cite them, with any namesakes offered alongside so you can pick the right one. If nothing matches every word, it loosens the search rather than showing you an empty page.
A plain search looks at titles and abstracts. That is a deliberate limit: searching the full text of every paper sounds better and is in practice much worse, because a word in a reference list counts as a match and most of what comes back is then about something else. On one measured search it was the difference between 2,159 results and 152. The wider search is still there for when you want it.
An API key is a password that lets a program use an AI account in your name. You create one at OpenAI or Anthropic and pay them directly for what you use, which for a single paper is usually a few cents. Nothing is resold here and there is no markup, because the key is never mine to bill against.
You can also skip the key entirely and choose Local LLM, which points the site at a model running on your own computer through Ollama, LM Studio or vLLM. You give it an address and a model name. No key, no cost, and the text of the paper never leaves your machine.
Which model does the quick work and which does the slow, careful work is read from a register that is kept up to date centrally, so a model that has been retired by its maker does not quietly keep being used months later. Everything except the analysis and translation features works with no key at all.
There is no analytics script, no advertising, no third-party tag and no consent banner, because there is nothing to consent to: what you search for is not stored against you, and your marks, notes and collections live in your own browser rather than on the server. Signed-in users get two buttons to push a copy to their account and pull it back, but only if you click. There is no background sync.
Two things are worth naming rather than glossing over. The web server keeps an ordinary access log — address, time, page, browser string — as every web server does, and I read it to see whether anyone is using the site at all. And the two typefaces on these pages are loaded from the Google font service, which is the one request that leaves this domain; serving them from here instead is on the list.
The search works and it is public. The honest summary: metadata coverage is about as broad as it gets, and full text is where it thins out. What has been fixed or added recently — including the things that turned out to be my own mistakes — is listed on the what changed page.
A long list of papers is not an answer to a question. Beside the results there is a panel counting who wrote them, what they are about, when they appeared and how much is open access, over the whole search rather than the page in front of you. It tells you which few people carry a field, and every line doubles as a filter: keep your search, narrow it to one author, one subject or one year, click again to undo.
Every result carries quality indicators, so you see at a glance how often a paper is cited, how recent it is and whether it even has an abstract. From there you can compare papers side by side, collect them in a dossier, follow the citations, translate the text, or have it analysed, including a deliberately adversarial reading that argues against the paper rather than summarising it.
The interface runs in Arabic, Dutch, English, French, German, Portuguese, Spanish and Turkish, and any paper can be translated on the spot into considerably more, with an honest label per language: an AI is better at some than at others, and the site says which rather than pretending they are equal. Open access is about who can reach the research, and language is one of the harder walls in the way.
What people need is the best relevant research, not only the part that happens to be free. Records behind a paywall appear alongside open ones, marked as such and with the price where it is known, so you can see what a paper would cost before you decide. One switch at the top of the results limits everything to open access, so the choice is yours rather than mine.
Nothing is stored on my side by default. Searches, marks, notes and collections live in your own browser. Signed-in users get two buttons to push a copy to their account and pull it back, but only if you click. There is no background sync.
The analysis and translation features run on an AI key you supply yourself, so the cost and the choice of model stay with you. It lives encrypted in your browser and is deliberately excluded from the account sync, so it never lands in any database of mine. When you ask for an analysis the key does pass through this server on its way to the AI provider: held in memory for that one request, never stored, never written to a log. Everything else, including all searching, works without a key.
Metadata is the easy part. Getting hold of the article is not: roughly one in four open-access papers here still fails to hand over its text, because the publisher blocks automated access or the only copy is a scan. Where public money required a free copy — NIH-funded medicine in PubMed Central, for instance — even paywalled papers are usually within reach; where it did not, they usually are not. CORE, which aggregates repositories worldwide, is already used to find a copy; its own full text needs a free non-commercial key that is not yet in place, and that is the next step.
The Smithsonian sits alongside OpenAlex because I connected it by hand. Any archive or repository that wants the same should not have to email a stranger and hope, so there should be a straightforward route to put itself forward.
OpenAlex, the catalogue this runs on, moved to usage-based pricing in February 2026. A free key covers roughly a thousand searches a day — more than this site currently uses, though a single researcher in a real session can easily spend twenty of them. It belongs on this page because it is the same pattern open access set out to break, moved one layer down: not the papers, but the catalogue of them. If this site ever outgrows the free allowance the cost lands on me, which makes a second source less a feature than an insurance policy.
Everything currently runs on one machine in one country. That is a single point of failure dressed up as independence. Mirrors elsewhere need time and allies, so this is a direction rather than a plan.
Open access has succeeded at scale, and that success created a second problem. There are hundreds of millions of papers within reach, more than anyone can read and far more than anyone can judge. Search engines are good at handing you a thousand results and silent about which of them deserve your time.
So the aim is narrow: not to help you find more, but to help you decide faster. Show what a paper costs, how often others leaned on it, how old it is, how it reads next to the one beside it, and what someone determined to disagree with it would say.
No. Searching, filtering, comparing and reading work without signing in. An account exists only so that a collection you built can be carried to another device, and it is created by signing in with an ORCID iD.
Yes. There is no paid tier and nothing is held back for one. The only thing that can cost money is the optional AI analysis, and you pay your own AI provider directly for that, not this site.
The site has no AI account of its own. If you want summaries, translations or an analysis, you supply an API key from OpenAI or Anthropic, or point the site at a model running on your own computer. You choose the model and you see the bill.
It is stored encrypted in your own browser and is deliberately left out of the account sync, so it never reaches any database. When you ask for an analysis the key passes through the server on its way to your AI provider: held in memory for that one request, never stored, never written to a log.
Yes, and most of the site is built that way. Searching, the result panel, the filters, the quality indicators, side-by-side comparison, citation chains and collections all work with no key present.
OpenAlex supplies the index, which is where records from arXiv, PubMed, Zenodo, bioRxiv and institutional repositories are reached. Unpaywall, Europe PMC and CORE are used to find a free copy of a paper, ORCID to pull in your own work, and Smithsonian collection records are searched alongside the articles.
There is no analytics script, no advertising, no cookie asking for consent and no profile. Searches are not tied to a person. The web server keeps an ordinary access log, as every web server does, and the two web fonts are loaded from the Google font service.
Models from OpenAI or Anthropic with your own key, or any model you run locally through Ollama, LM Studio or vLLM. Which model is used for a quick task and which for a deep one is read from a register that is updated centrally, so a retired model does not silently keep being used.
On a machine owned and run by one person in Antwerp, Belgium. There is no cloud platform in between, no company and no investors.
No, the code is not published. What the site does, what changed and what went wrong is written down in plain language on the what changed page instead.
Written by me with AI assistance, which is worth stating plainly rather than leaving for someone to notice. The judgement calls, the sources and the mistakes are mine.