A tiny reference manager backed by a Dropbox folder of PDFs and published as a static web page via GitHub Pages.

Live page: https://timodonnell.github.io/papersimreading/

A scheduled job scans a Dropbox folder for new PDFs, extracts each paper's metadata (title, authors, journal, publication date, abstract, and a link to the PDF) from public sources, and appends a record to data/references.json. The web page (index.html) renders that JSON — searchable and sortable, no build step.

Only real papers are published. A file is included only if it resolves to a public identifier (a DOI or arXiv id). Anything else — books, grants, manuals, reviewer copies, personal files — is skipped entirely and never linked. The list of skipped files is remembered locally in a git-ignored .papersync-excluded.json so those filenames never reach the repo and are not re-processed each run.

Link policy. Open-access and arXiv papers link to the free public copy (resolved via Unpaywall). Paywalled papers (a DOI, but no open copy) link to the user's own PDF in Dropbox. So a Dropbox share link is only ever generated for an identified paper — never for an unrecognized file. When an HTML edition exists (arXiv native HTML or a PubMed Central full-text page), it is linked too.

Dropbox folder (synced locally)
        │  scan for *.pdf not yet recorded
        ▼
  papersync.sync ──► extract DOI / arXiv id / title from the PDF
        │            resolve metadata:
        │              1. Crossref by DOI
        │              2. arXiv API by arXiv id
        │              3. Crossref by filename-derived DOI  (bioRxiv/Nature ids)
        │              4. Crossref by title
        │              5. LLM over first-page text  (optional, needs API key)
        │            no DOI/arXiv id? -> not a paper, excluded (never published)
        │            link: OA/arXiv -> public copy; paywalled -> Dropbox
        ▼
  data/references.json ──► git commit & push ──► GitHub Pages renders index.html

The Dropbox folder is synced to this machine by the Dropbox desktop client, so the job reads the local copy directly — no scraping, no API auth, no bulk downloads. The shared-folder URL and folder path are never committed; they live in a local config file.

1. Install dependencies (poppler provides pdftotext):

   pip install -r requirements.txt
   sudo apt-get install poppler-utils    # if pdftotext is missing

2. Create the local config (outside the repo):

   mkdir -p ~/.config/papersimreading
   cat > ~/.config/papersimreading/config.json <<'JSON'
   {
     "papers_dir": "/home/you/Dropbox/your-folder",
     "crossref_mailto": "[email protected]",
     "anthropic_api_key": "",
     "generate_dropbox_links": true
   }
   JSON
   chmod 600 ~/.config/papersimreading/config.json

   Set anthropic_api_key to enable the LLM fallback for PDFs that no public source can identify. Leave it empty to stay API-only (no token cost).

python3 -m papersync.sync              # process every new PDF
python3 -m papersync.sync --limit 20   # stop after 20 new papers
python3 -m papersync.sync --since 90   # only PDFs modified in the last 90 days
python3 -m papersync.sync --dry-run    # list new PDFs, write nothing

The run is incremental and resumable: records are flushed to disk every few papers, and already-processed files are skipped next time. Files are tracked by content hash, so renaming or moving a PDF within the folder does not re-process it.

scripts/run.sh runs the sync and, if references.json changed, commits and pushes. Install it as a daily cron job:

scripts/install-cron.sh                       # daily at 07:30
CRON_SCHEDULE="0 * * * *" scripts/install-cron.sh   # hourly

Because the PDFs live in this machine's local Dropbox sync, the job must run on this machine (not a cloud agent). Logs go to papersimreading.log (git-ignored).