Dear Paperpile staff and users,
One strength of paperpile is not only a reference manager, but a massive library of pdfs that agents can consult when running research. It makes more sense to use your paperpile library as the central pdf repository rather than having agents try to download pdfs into separate folders each time. However when agents find papers online that are missing from my library, I needed a tool to quickly add them to the library. So I wrote a quick repo that brings up a Pubmed page of the missing pdfs, so you can quickly add them to your library.
I’m not sure about the status of MCP servers and whether letting agents click the equivalent “add to paperpile” or “add pdf” buttons would trigger bot detection measures.
Finally, I appreciate your work getting around the annoying “blocked by Captcha” issue of which ScienceDirect is the primary culprit.
Best regards,
David Priest
AI summary below.
I do a lot of literature work with AI coding assistants now, and hit a problem that I suspect is about to become common: they can search PubMed perfectly well, but they cannot see the 2,100 PDFs already sitting in my Paperpile folder.
It came to a head recently. I had an agent verify an argument against twelve primary papers. It searched, found them all online, and had me download eleven by hand. Every single one was already in my library. Nobody was at fault — the agent had no way to look, and I didn’t think to check.
So I wrote three small tools to close that loop: GitHub - david-priest/paperkit: Find, fetch and gap-check a local reference library from the command line · GitHub — MIT, unofficial, no affiliation with the Paperpile team.
paperfind Bannard 2013 # is it already on disk?
paperget "Chou 2016 AP4 chronic viral" # it isn't — one click to the PDF
papergap --pdfs ./refs --gaps # what is this reading list missing?
The workflow it enables. Hand an agent a reading list. It runs paperfind on each, tells you which ones you actually lack, and paperget turns the remainder into a single page of links. You add them; the agent reads the ones you already had in full, rather than working from abstracts.
A trap worth knowing about even if you never install this. My synced Paperpile/ folder has only ~62 PDFs in it. After a sync problem last year, the real library was left behind as a stale local copy somewhere else. Anything searching the obvious location concludes a paper is missing when it is right there. If you have ever had a sync hiccup, it is worth checking whether you have more than one copy of your library in different states.
The bit that surprised me. I wanted a page of download links with a Paperpile button on each row. That does not work on a local HTML file — the extension only injects on sites it carries a parser for. But it works fine on a PubMed results list. So an OR-query over PMIDs gives you one tab, every paper you want, one Paperpile button each:
https://pubmed.ncbi.nlm.nih.gov/?term=12021782+OR+15300245+OR+27566939
Useful for bulk-adding a reference list by hand, with no tooling at all. One caveat found the hard way: adding a batch this way will rate-limit Paperpile’s own PDF fetch (menu_pdf.RATE_LIMITED). Adding the references still works; for the PDFs, going straight to the PMC copy avoids the issue.
Caveats. Built against my own library, so the filename parsing assumes Paperpile’s <Surname> et al. <Year> - <Title>.pdf convention. macOS and Linux, Python 3, nothing to install. paperfind -m reads the local Realm metadata cache read-only, which is handy for recovering a DOI when a PDF has gone missing.
The matching is fussier than you would expect, because every looser version confidently told me I owned a paper I did not. At one point “Shih 2002” matched an Optics Letters study of bismuth thin films — genuinely from 2002, genuinely with an author named Shih. Only the title separates them. Both resolvers carry regression suites of exactly these cases, because a confident wrong answer is worse than no answer.
Would be glad to hear if it breaks on someone else’s setup, particularly anyone whose library is not on Google Drive — I have not been able to test that.