eprint

Fast command-line search over the IACR Cryptology ePrint Archive.

Metadata for all ~26,000 papers is harvested from the archive's OAI-PMH interface into a local SQLite FTS5 index, so searching is instant and works offline.

$ eprint "threshold ecdsa" -n 3

3 of 62 results  in: title, authors, abstract · index 8m old

  2026/1455    Trout++: Robust Asynchronous Two-Round ECDSA for Arbitrary Thresholds
               Nof, Parker

  2026/1103    Jevil: A Catastrophic-Failure-by-Design Signature Scheme
               Kobeissi

  2026/976     Revisiting DKLs Threshold ECDSA: Enhanced OT-based VOLE and Two-Party Signing
               Asharov

Results are titles and authors, newest first — the second hit above matches in its abstract rather than its title. Add -a for full abstracts, -t to match titles and authors only, or eprint show <id> for one paper in full. For reading rather than looking something up, there is a full-screen interactive browser.

It can also keep up on its own: desktop notifications when a paper matching one of your watches appears, and the browser one ⌘Space away.

Opening papers

On terminals that implement OSC 8 hyperlinks (iTerm2, Ghostty, kitty, WezTerm, VS Code, recent GNOME Terminal), paper ids and titles are themselves clickable.

macOS Terminal.app does not support OSC 8 — it only auto-detects plain text URLs. It is detected automatically and a bare, cmd-clickable URL is printed under each result instead. Force either behaviour with --urls / --no-urls.

eprint open <id> works everywhere regardless of terminal.

Papers you open are kept

eprint open <id> goes straight to the PDF — the landing page is a detour, and eprint show already holds the metadata it would have given you. Opening a paper is also taken as a signal that you want it, so the tool quietly builds a local library at ~/Documents/eprint/:

• First open — the PDF opens in your browser:

  $ eprint open 1523
  ⌘S anywhere in Downloads, Desktop or your home folder and it will be kept
  

  Press ⌘S (browsers display PDFs inline rather than downloading them) and accept whatever folder the dialog suggests — no navigating. Downloads, Desktop, your home folder and the library itself are all watched, because a browser suggests wherever it last saved and no command-line tool can change that. The file is filed as 2026-1523-catching-many-traitors.pdf.

• Every open after that — the local PDF opens directly. Instant, offline, no browser.

The same applies to enter in browse, which shows the same hint in its status line. There are no flags and no settings for this; the only visible difference is that the second open is instant. Set EPRINT_PAPERS_DIR or EPRINT_DOWNLOAD_DIR if the defaults are wrong for you.

Filing is deliberately conservative. A download in a watched folder is only adopted if it is named for the paper (ePrint serves /YYYY/NNNN.pdf, so browsers save 1523.pdf), appeared after you opened the paper, is not a partial download, starts with the %PDF- magic bytes, and has stopped changing size. It is copied, not moved, so nothing vanishes from where you saved it. EPRINT_DOWNLOAD_DIR replaces the watched set if your browser saves somewhere unusual.

A PDF saved into the library folder is treated as deliberate: it is renamed to the canonical name and used, whenever it arrives — so saving there works even hours later, long after the watcher has exited. A file whose name claims a different paper (2025-1523-…) is never served for 2026/1523, and a file that is not really a PDF is ignored.

Clicking an OSC 8 link in your terminal bypasses the tool entirely, so those opens cannot be cached — use eprint open or browse if you want the local copy.

To drop a paper you are done with:

eprint open --rm 1499              # or several: --rm 1499 2026/1464

It names each file it deletes and reports what was freed, and it only ever touches the library folder. Nothing is lost for good — opening the paper again fetches it back. The filenames are canonical and readable precisely so that rm ~/Documents/eprint/2026-1499-*.pdf works just as well if you would rather not go through the tool.

The tool never downloads a PDF itself, by design. The archive serves them behind a Cloudflare challenge, and its robots.txt says "Full text PDFs are only available under a license specific to each paper" while denying *pdf to every agent. Metadata is offered to machines over OAI-PMH; full text is not. So your browser does the fetching, and this only files what your browser saved.

Interactive browser

A full-screen reader for exploring the archive rather than looking one thing up.

eprint browse                      # everything, newest first
eprint browse "lattice signature"  # start from a query
eprint browse --author Boudgoust --date 2023..2024

Abstracts expand and collapse in place, matches stay highlighted inside the expanded text, and the query can be edited live with /. There is no cap on how much it loads — a bare eprint browse holds all ~26,000 papers and G really does reach the oldest one from 1996 — because only the rows on screen are ever laid out. -n still exists if you want a smaller set.

Key	Action
j / k, arrows	move
g / G, home / end	first / last
ctrl-d / ctrl-u, page keys	jump
space / tab	expand or collapse the abstract
t	toggle where the query is matched: in: title, authors, abstract ⇄ in: title, authors
d	filter by date — same grammar as --date; empty clears it
v	filter by venue — CRYPTO or CRYPTO 2025; prefixes work; empty clears it
w	show only papers matching a watch — searches apply within that subset
/	edit the query — results filter live as you type
ctrl-u	clear the query (while editing)
enter / o	open the PDF — your local copy once you have one
y	copy the URL to the clipboard
b	copy the CryptoBib citation key (published version when known)
B	copy the full BibTeX record
q / esc / ctrl-c	quit

/ refines the current query rather than replacing it, so you can narrow a search by typing more terms; ctrl-u clears it to start fresh. Expansion is tracked per paper id, so it survives re-searching. Typing is debounced: the query line updates on every keystroke, the listing catches up once you pause, and a … next to the cursor means it has not caught up yet.

Copying needs a clipboard tool on Linux

y, b and B shell out, because a terminal cannot write your clipboard by itself. macOS has pbcopy built in; Linux does not ship anything by default, so install one:

sudo apt install wl-clipboard      # Wayland — the Ubuntu default session
sudo apt install xclip             # X11

wl-copy, xclip and xsel are all tried, in whichever order suits the session, so either package is enough and having both is fine. If none is present the status line says so and names the package rather than just reporting failure.

Terminals that implement OSC 52 (kitty, WezTerm, foot, Alacritty, tmux, and so over ssh) are used as a last resort and need nothing installed. GNOME Terminal is not one of them — VTE has never implemented it — which is why the packages above matter there.

Because nothing needs to be clicked, browse works identically on terminals without OSC 8 support — press enter instead.

Install

Requires a Rust toolchain (1.86 or newer). If you do not have one: https://rustup.rs

git clone https://github.com/msimkin/eprint.git
cd eprint
cargo install --path . --locked

That puts an eprint binary in ~/.cargo/bin, which rustup already adds to your PATH, so you can then run eprint from anywhere.

Use --locked. cargo install ignores Cargo.lock unless you pass it. Without it Cargo re-resolves dependencies and pulls versions of darling and instability that require rustc 1.88, so the build fails on 1.86. The committed lockfile pins compatible versions. On rustc 1.88 or newer you can drop the flag, or run cargo update to lift the pins.

Check it worked:

eprint --version

The first search builds the local index automatically — about 30 seconds, once.

Switch on Tab completion while you are here (zsh or bash):

eprint config --completions

Without installing

To run it from the build directory instead:

cargo build --release --locked
./target/release/eprint --help

If eprint is not found

~/.cargo/bin is missing from your PATH. Add it to your shell's startup file:

# ~/.zshrc (zsh, the macOS default) or ~/.bashrc (bash)
export PATH="$HOME/.cargo/bin:$PATH"

# fish
fish_add_path ~/.cargo/bin

Then open a new shell. Alternatively, symlink the binary somewhere already on your PATH:

ln -s "$PWD/target/release/eprint" /usr/local/bin/eprint

Updating and uninstalling

git pull && cargo install --path . --locked   # rebuild and replace
cargo uninstall eprint                        # remove the binary

Uninstalling leaves your index and config in place; see Storage to remove those too.

Use

eprint                                # papers that arrived since you last looked
eprint "threshold ecdsa"              # search
eprint "fully homomorphic" -a         # include full abstracts
eprint "Katharina Boudgoust" -t       # match titles and authors only
eprint garbled --date 2024 -n 50      # one year, more results
eprint --author Boudgoust --date 2y   # filter without a query
eprint --category "Public-key" -n 5   # filter by IACR category
eprint show 2026/1538                 # one paper in full
eprint open 2026/1538                 # the PDF — your local copy once you have one
eprint open                           # list the papers you have downloaded
eprint open --rm 1538                 # delete a downloaded copy
eprint watch add "lattice OR LWE"     # mark matching papers with a ✱ everywhere
eprint bib 2018/116 --entry           # the whole BibTeX record
eprint status                         # index stats
eprint update                         # refresh now
eprint update --full                  # rebuild from scratch
eprint update --quiet                 # refresh with no progress output
eprint config --completions           # switch on Tab completion (zsh or bash)

Every short flag has a long form: -n is --limit, -a is --abstracts, -t is --title.

A query is just the first argument, so eprint search <query> and eprint <query> are the same thing; search is kept for the fingers that expect it. new and Bib are not — those words are now ordinary query terms.

Tab completion (zsh and bash)

eprint config --completions      # adds one line to your shell's rc file

It reads $SHELL, writes to ~/.zshrc or ~/.bashrc accordingly, is idempotent, tells you the file it touched, and does nothing if the line is already there. eprint config reports whether completion is on. To do it by hand instead:

echo 'eval "$(eprint completions zsh)"'  >> ~/.zshrc
echo 'eval "$(eprint completions bash)"' >> ~/.bashrc

Either way that one line is enough even on a bare rc file: the function ships inside the binary. Under zsh it initialises the completion system itself if your shell has not already done so — without that, compdef is undefined and Tab does nothing anywhere, not just for eprint, which is a confusing way to discover that your shell never ran compinit. The bash function needs no such bootstrap, and deliberately depends on nothing outside bash itself: not the bash-completion package, and no syntax newer than bash 3.2.

The two are not quite equal. Bash has no description column, so where zsh shows a paper's title or an author's paper count beside each candidate, bash can only offer the value. Everything else — the commands, the per-command flags, paper ids, categories, authors, watch numbers, and the --flag=value form — completes the same in both, case-insensitively in both.

eprint open <TAB> then offers the papers you have already downloaded, with their titles:

$ eprint open <TAB>
2026/1464  -- Optimal Distributed Monotone-Policy Encryption for DNFs and More from Lattices
2026/1499  -- BF²: A Bloom-Filtered Brute-Force Framework for Multi-Target Password Recovery
2026/1523  -- Catching Many Traitors in Threshold Traitor Tracing: Lower Bounds and Constructions

show and bib complete the same set, eprint <TAB> the commands, eprint watch <TAB> its three verbs. Typing an id that is not in the library still works — it just opens online, as always.

--category <TAB> offers the archive's categories — there are only seven, nobody remembers their exact wording, and they are the one filter you cannot guess:

$ eprint watch add --category <TAB>
Applications               -- 2030 papers
Attacks and cryptanalysis  -- 1281 papers
Cryptographic protocols    -- 6224 papers
Foundations                -- 3090 papers
Implementation             -- 2364 papers
Public-key cryptography    -- 4792 papers
Secret-key cryptography    -- 2885 papers

The list is read from your index rather than baked into the binary, so a category the archive adds shows up without a new release, and the counts tell you whether a filter is worth having. Any substring works too — --category proto is the same filter — and a name with a space is quoted for you as you complete it. This works wherever --category does: searches, browse and watch add.

--author <TAB> completes names out of the index once you have typed a few letters — the full list is 19,540 names, so it narrows rather than dumps. Names are offered in both orders, since completion matches on a prefix and you may start from either end:

$ eprint watch add --author boudg<TAB>
Boudgoust, Katharina  -- 18 papers
Boudguiga, Aymen      -- 13 papers

$ eprint watch add --author Katharina\ B<TAB>      → Katharina\ Boudgoust

Boudgoust, Katharina and Katharina Boudgoust are the same filter — the comma is there only so the candidate starts with what you typed, and the space is escaped for you. Candidates are always people, never bare surnames, and spellings that differ by accents, punctuation or spacing are shown as one person with their papers added together.

eprint watch rm <TAB> offers your saved watches by number, with what each one is:

$ eprint watch rm <TAB>
1  -- by Boudgoust
2  -- lattice OR LWE
3  -- in Foundations

Those numbers are positions that renumber after every removal, so having them listed beats counting by hand. --scope and --theme complete their values too, and flag names complete as well — eprint lattice -<TAB> lists what a search takes, eprint bib -<TAB> what bib takes. Both --category <value> and --category=<value> are understood.

Completion ignores case everywhere, like the filters themselves: --author shamir<TAB> and --category crypto<TAB> work as well as the capitalised forms. Names insert with the archive's own capitalisation but without accents, since a shell cannot reach Damgård from a typed damga; the description beside the candidate shows the real spelling when the two differ.

Changing completion means starting a new shell, or exec $SHELL — the function is loaded once when the shell starts, so an already-open terminal keeps the version it read at login.

Paper ids complete from your library, not the whole archive: a short changing list is what completion is good at, and 26,000 candidates would mean megabytes of output per keypress. --author is the one target that is filtered rather than listed, for the same reason — the full set is 19,540 names — so it answers once you have typed two letters. Nothing completes for --date or for watch query terms, where the candidate set is "anything you might type".

Without any shell setup, eprint open with no id lists the same thing:

$ eprint open
  2026/1523    1.2 MB  Catching Many Traitors in Threshold Traitor Tracing: Lower Bounds and Constructions
  2026/1522    0.4 MB  Efficient Ternary Computation of Optimal Ate Pairing on BLS27 Curves
  …
  6 papers, 5.5 MB in /Users/you/Documents/eprint

The sizes are there because "which of these can go?" is the question that comes before eprint open --rm.

Desktop notifications and a launcher

Two optional pieces of desktop integration, both off until you ask, both removed the same way.

Notifications for new papers

eprint config --notify summary   # off | all | summary | watched
eprint config --notify off       # and it is gone again

That writes the mode to the config file and installs a background updater — a launchd agent on macOS, a systemd user timer on Linux — which harvests every 30 minutes and announces what arrived. A test banner is posted straight away, and the output says which sender posted it — read that line, because it also spells out whatever setup step is still missing.

mode	what you get
off	nothing. The default
summary	one banner: 7 new papers
all	one banner per paper, capped at five, then +N more
watched	only papers matching a watch, each naming the watch it matched

ePrint posts in bursts of forty, so watched is the mode worth having. A harvest you asked for never posts a banner; one that happened behind your back does. eprint config and eprint status both say whether the updater is installed.

Banners use whatever the platform already ships. On a stock macOS that is osascript, which works but poorly: the banner is attributed to "Script Editor", and clicking it opens Script Editor rather than anything useful. Two steps fix both:

brew install terminal-notifier
eprint config --notify summary    # run again, now that it has something to wrap

With terminal-notifier present, config --notify wraps its binary in a small eprint.app (kept under ~/Library/Application Support/eprint/), so banners carry eprint's own name and icon — and clicks work: the summary banner opens eprint.iacr.org, a per-paper banner opens that paper's page. One manual step remains: open System Settings → Notifications, find eprint in the list, and switch Allow notifications on. macOS never shows its usual permission prompt for this bundle — it is assembled on your machine and only ad-hoc signed, and macOS silently declines such apps instead of asking — so until the switch is flipped, banners fall back to the best sender that is allowed, and the config --notify output names it. The toggle is needed once; it survives upgrades and reinstalls, and eprint config --notify off removes the bundle along with the updater.

Linux needs notify-send (sudo apt install libnotify-bin); banners post under eprint's name as they are. There is no click-through there by design: notify-send can only observe a click by blocking until the banner is dismissed, and a scheduled job must never hang on one.

Opening browse from Spotlight

eprint config --launcher on

On macOS this writes ~/Applications/eprint.app. Press ⌘Space, type eprint, hit ⏎: a Terminal window opens, updates the index in front of you, and goes into eprint browse. Quit with q and the shell is still there, so you are left at a working prompt. On Linux it writes ~/.local/share/applications/eprint.desktop, which your desktop's application search finds and opens in whatever terminal you already use.

Spotlight will offer "press Tab to search". Ignore it — that is macOS's app-scoped search, which eprint has no index to answer; press ⏎ without Tab. macOS offers it for every application.

macOS defaults to Terminal.app. For anything else set terminal_command, with {cmd} standing in for the command to run:

terminal_command = ghostty -e {cmd}

eprint config --launcher off removes it, and refuses to delete a bundle it did not write.

Keeping up

eprint                         # arrivals since you last ran it, never empty
eprint -n 20                   # exactly 20, batch or not

A bare eprint — no query, no filters — is the feed rather than a search, and it always shows something:

• More arrived than latest_limit? You get the whole batch. Seventeen papers means seventeen lines, not a number chosen in advance.
• Fewer, or nothing new? The list is topped up with the most recent arrivals to latest_limit (10 by default), and the header says how many are actually new — 2 new since 31/07/2026, or nothing new since 27/07/2026.
• -n N overrides both as an exact count.

So latest_limit is a floor, not a ceiling.

The two dates mean different things, on purpose. 2 new since 31/07/2026 counts what arrived since you last looked, because that window is what the number refers to. nothing new since 27/07/2026 instead reports when the archive last posted — dating that line by your last run would only ever tell you that you ran the command recently. When the index itself is more than a day old, the header says so (· index 3d old), because a stale index otherwise looks exactly like a quiet archive.

Papers carry an added timestamp recording when they first entered your index, which is what this filters on. A paper's own date can predate its arrival here, so filtering by that would silently skip late-published submissions.

The index refreshes in the background, the same as for a search, so this stays instant even when the local copy is stale. The batch replay below is what makes that safe: a slightly old answer is shown again next time rather than lost.

The last batch sticks around. ePrint posts in bursts, so most runs find nothing at all. Rather than print "nothing new" for the rest of the day, the feed shows the last batch it found again, and says so:

4 results  last batch, from 30/07/2026 · nothing new yet

When the archive does move, you get only the genuinely new papers and those become the remembered batch — so each burst is shown until the next one replaces it, and nothing is ever silently skipped.

Watches

A watch is a saved search that marks the papers you care about. It is purely a highlighting feature: it never changes which papers are shown, how many, or in what order.

eprint watch add "lattice OR LWE"            # papers mentioning either term
eprint watch add --author Boudgoust          # everything by one author
eprint watch add --author "Katharina Boudgoust"   # a full name needs quoting
eprint watch add zk --category "Public-key"  # "zk" AND in that IACR category
eprint watch add "proof of work" -t          # those words in the title or authors
eprint watch                                 # list them, numbered
eprint watch rm 2                            # remove one
eprint watch rm --all

eprint watch lists them the way you would say them, with how many papers each one currently marks — the quick check that a new watch matches anything at all:

$ eprint watch
  1   lattice OR LWE                               2681 in the index
  2   by Katharina Boudgoust                       18 in the index
  3   zk · in Public-key cryptography              36 in the index
  4   proof of work · titles only                  56 in the index

An author filter matches every word of the name, in any order, ignoring accents, punctuation and case. --author "Katharina Boudgoust" and --author "Boudgoust Katharina" are the same filter, and so are Damgård and Damgard — the archive contains both spellings of the same person, along with Ron D. Rothblum and Ron D. Rothblum differing only by a stray space. Single letters are dropped, which is what makes Ron D. Rothblum and Ron Rothblum one filter rather than two. A full name has to be quoted, or the shell hands Boudgoust to the query instead of to --author — which is what Tab completion is for (below): it fills in the whole name and escapes the space for you.

One author, however the archive spells them. The same person is often filed several ways, and the tool settles on one spelling for them everywhere — in a listing, in show, as a completion candidate and in a watch count.

Two things do that. First, a fold: accents, punctuation, case, stray spaces and the written-out form of an accented letter (ö/oe, å/aa, ü/ue) all come to the same thing, so Damgaard, Damgard and Damgård are one filter, and so are Doettling and Döttling. It is careful about where it does that — ue only inside a word — because Yu and Yue, or Xu and Xue, are different people, and every case that needs the rule (Gueneysu, Kuesters, Buenz, Mueller) has the digraph mid-word.

Second, a table of 428 well-published authors — everyone with ten or more papers whom the archive spells more than one way — built into the binary, for the rest: a bare initial (N. P. Smart, F. Vercauteren), a hyphen splitting a given name (Hwa-Jeong Seo), a middle name that comes and goes (Yael Tauman Kalai, Ron D. Rothblum), one typo, and one paper where the archive put two people in a single author field. Names are written through it as they are harvested, so:

Damgard, Ivan    -- 142 papers
Damgard, Kasper  --   1 paper

rather than five entries for Ivan Damgård and one for Kasper. There is deliberately no rule that guesses this: expanding S. Sree Vivek or T-H. Hubert Chan to the archive's commonest name for that surname produces a different person, so the corrections are hand-checked data rather than a heuristic, and an author with one or two papers and an unusual spelling may simply be missed. A name is also matched against one author at a time: --author "Kasper Damgård" does not find papers where Kasper Larsen and Ivan Damgård happen to be co-authors.

Adding a watch you already have is not an error; it says so and changes nothing.

Matching papers get a gold ✱ after the title and their id in the same gold, everywhere a list of papers appears — searches, the feed and browse:

  2026/833     Scale, Round, Break: Simple Leakage Attacks on Secret Sharing Schemes ✱
               Boudgoust, Simkin

  2026/459     Naor-Yung Transform for IND-CCA Probing Security with Lattice Instantiations ✱
               Boudgoust, Imbert, et al.

The badge follows the title text rather than sitting in a fixed column, so only a watched row gives up any width for it — its title wraps two columns narrower, leaving room on the last line. On a wrapped title the badge lands at the end of the final line. The glyph is U+2731 HEAVY ASTERISK, single-width in every terminal where ★ and ◆ are East-Asian-Ambiguous and could overrun the line. It degrades to a plain ✱ under NO_COLOR or theme = "mono".

The gold is 256-colour index 136 (#af8700), the one place this tool does not use a plain 16-colour index: "a shade darker than the match highlight" is not something a palette index can promise, since a theme's yellow could be anything.

In browse, w filters the listing down to watched papers, and any query you then type searches within that subset. The header shows · watched only while it is on; w again restores everything.

Which papers match is worked out once and kept in the index, along with which watch matched them, so toggling w, typing in the query box, and the counts above all read from a table rather than searching the archive again. Adding a watch matches only that watch; removing one is a delete. Only new papers arriving cause everything to be re-matched, and that happens in the background refresh — never on the command you are waiting for.

Watches live in the config file, one watch line each, written exactly as you would type them:

watch = lattice OR LWE
watch = --author Boudgoust
watch = zk --category Public-key   # terms and filters combine: both must hold
watch = proof of work --title

Values are written the way the tool writes them, and quoting means the same thing it means on the command line: bare words are separate terms, "a b" is an exact phrase. So watch = lattice OR LWE is an OR, while watch = "lattice OR LWE" would look for that literal phrase and match nothing.

An author is saved as a name reads, whichever candidate you picked: completing Shamir, Adi stores --author "Adi Shamir", and a name the table knows is saved as its usual spelling, so a list of two dozen watches does not have one odd entry in it.

So copying ~/.config/eprint/config.toml to another machine copies your whole setup — theme, limits and watches together — and you can add or remove them by hand with eprint config --edit as easily as with watch add. watch add/watch rm rewrite just those lines and leave the rest of the file, comments included, untouched. Watch numbers are positions in the file, so they close up after a removal.

Watches store no year filter — that would date a standing watch — and the count eprint watch shows next to each one is its total across the whole index, which is the quick check that a new expression actually matches something.

If you used an earlier version, your watches were kept in the index database; the first run of a newer build moves them into the config for you and says so.

Citation keys (CryptoBib)

It refreshes itself weekly. Nothing below is needed for ordinary use: the background update that keeps the index current also asks CryptoBib, at most once a week, whether it has published anything new. That check is a conditional request costing nothing when the answer is no, and CryptoBib publishes a handful of times a year. eprint bib reports both dates — when the data was last published and when it was last checked — because a months-old publication date beside a recent check is a quiet upstream, not a stale index.

eprint bib --update          # download / refresh the CryptoBib database
eprint bib --update --force  # re-download even if unchanged
eprint bib 2018/116          # citation key only
eprint bib 2018/116 --entry  # the whole BibTeX record
eprint bib                   # database status

--entry prints the complete record rather than just the key — the command-line equivalent of B in the interactive browser, where b copies the key and B the whole entry.

Links each paper to its CryptoBib citation key, preferring the published version over the preprint where one is known:

$ eprint bib 2018/116

  EC:CGKW18
  published version

In browse, b copies the citation key and B copies the complete BibTeX record. The key is shown in the meta line once an abstract is expanded, and eprint show displays it too.

Where a paper was published

The same data answers a more useful question, and it is shown without asking:

  2024/1291    Raccoon: A Masking-Friendly Signature Proven in the Probing Model
               Pino, Katsumata, et al. · CRYPTO 2024

About 42% of the archive carries a venue — everything CryptoBib has matched to an ePrint paper. It appears on the byline in listings, in browse, in eprint show, and as a venue key in --json (absent, not null, when there is none).

--venue filters on it, in searches and in browse:

eprint --venue CRYPTO                # every CRYPTO paper
eprint --venue "CRYPTO 2025"         # one year
eprint --venue euroc lattice         # prefixes and case do not matter
eprint zero knowledge --venue TCC -a

One flag carries both parts, the way --date carries both ends of a range. A name is matched case-insensitively and then by unique prefix, and a name that matches nothing is an error rather than an empty listing — an empty listing would read as "nothing was published there". Tab completion offers the venues in the index, flagships first. In browse, v opens the same prompt.

Venue names are the series without the edition: CRYPTO, not Advances in Cryptology — CRYPTO 2024, Part III. Journals are spelled out and the IACR prefix dropped, since the archive is IACR's and saying so adds nothing — Transactions on Symmetric Cryptology, not IACR Trans. Symm. Cryptol..

A venue is not something you can watch. It arrives long after the ePrint posting, so a venue watch would almost never fire on a new arrival, which is the whole job of a watch.

eprint bib 2018/116 --entry >> refs.bib

Copied records are self-contained. Entries in crypto.bib reference venue names, publishers and editors as @String macros defined in a separate abbrev3.bib, so a record taken verbatim would not compile on its own. --update fetches both files and inlines the definitions (resolving nested references and # concatenation). BibTeX's built-in month macros are deliberately left unbraced so styles still render them properly.

Staleness. CryptoBib is never refreshed automatically. Once the local copy is more than 30 days old, browse shows bib 45d old in its header, copy confirmations append a reminder, and eprint bib prints a note.

Coverage. Of 25,692 linked papers, 10,678 (42%) resolve to a published version; the rest fall back to their EPRINT: key, and papers absent from CryptoBib entirely fall back to cryptoeprint:YYYY/NNN.

Matching is by exact normalised title plus at least one shared author surname. The author check matters: without it, unrelated papers sharing a title get linked. Fuzzy title matching was measured and rejected — it lifts coverage only from 42.8% to ~47.6% while introducing false positives. The residual misses are mostly papers whose title changed between preprint and publication, or whose published version is not in CryptoBib.

Refreshing is cheap. The server supports ETag, so an unchanged database returns 304 with an empty body and costs well under a second. A full rebuild downloads ~41 MB and takes about 70 seconds. There is no automatic refresh — run --update when you want it.

CryptoBib is fetched at runtime and cached locally in your own database; nothing is redistributed with this tool.

Configuration

eprint config          # show the file location and effective settings
eprint config --edit   # open it in $EDITOR, creating it first if needed
eprint config --init   # write a commented default file without opening it

Lives at ~/.config/eprint/config.toml (override with $EPRINT_CONFIG):

theme = "auto"       # auto | dark | light | mono
scope = "all"        # all | title
limit = 20           # results for a search
latest_limit = 10    # fewest shown by a bare `eprint`
notify = "off"       # off | all | summary | watched
watch = --author Boudgoust   # zero or more; see Watches below

Despite the .toml name the format is a plain key = value reader — no tables, no arrays, and watch may repeat, which real TOML forbids. Values may be quoted or bare.

limit caps a search. latest_limit is the floor under a bare eprint — see Keeping up — and applies only when there is no query and no filter; any query term or filter (--author, --date, --category) makes it a search. -n overrides both.

Results list authors only. The date joins them once an abstract is open (space in browse, -a inline). Category and licence are shown by eprint show.

Set notify with eprint config --notify <mode> rather than by hand — that also installs the background updater the notifications come from.

Command-line flags override the config file, which overrides the built-in defaults.

Colour

A muted 256-colour palette — brass and verdigris — picked so that the watch badge is the loudest thing on screen:

	dark	light
paper ids	66 #5f8787	30 #008787
URLs (underlined)	103 #8787af	61 #5f5faf
authors, dates, footers	102 #878787	240 #585858
watch badge	136 #af8700	136 #af8700
query matches	black on bright yellow	black on bright yellow

Matches set both foreground and background, so highlighting keeps its contrast on either ground. mono drops colour entirely and uses only bold, dim, underline and reverse.

These are fixed indices rather than your terminal's own first sixteen colours, which is a deliberate trade: the sixteen offer no muted mid-tones and their brightness is whatever your theme decides, so a palette built on them could not stay balanced. mono still inherits everything.

auto tries two things, cheapest first. COLORFGBG is free but only rxvt and Konsole set it; Terminal.app, iTerm2 and GNOME Terminal all leave it unset. So when it is missing, auto asks the terminal directly with an OSC 11 query and reads the rgb: background it answers with — the same mechanism vim, tmux and bat use for this. xterm, VTE (so GNOME Terminal), kitty, foot, WezTerm, Alacritty and Terminal.app all answer.

A terminal that stays silent costs one 100ms timeout and then falls back to assuming dark, as before. Setting theme explicitly skips the query entirely, and is still the right answer if you just want a particular palette.

Paging

Output longer than one screen is piped through a pager automatically. The default is less -RFX, which keeps colour, exits immediately when the output already fits, and leaves results in your scrollback instead of clearing the screen on exit.

Override with $EPRINT_PAGER or $PAGER; disable with --no-pager.

Query syntax

Queries use SQLite FTS5 syntax with Porter stemming, so signature also matches signatures. Multiple bare words are ANDed.

Example	Meaning
lattice signature	both terms
"fully homomorphic"	exact phrase
lattice NOT signature	exclude a term
zk OR snark	either term
homomorph*	prefix match

Punctuation that FTS5 would reject (zero-knowledge, MPC (dishonest majority)!) is handled automatically — the query is retried with each term quoted.

Partial words match. The index matches whole tokens, so boud would not find Boudgoust on its own. Bare terms are therefore treated as prefixes automatically: boud behaves as boud*. Quoted phrases, operators, column filters and terms already ending in * are left exactly as written. Pass --exact for strict whole-word matching.

Dates

Dates are shown and accepted day-first: 28/04/2026. One flag, --date, carries a single date or a range, at whatever granularity you feel like:

Example	Meaning
--date 2024	all of 2024
--date 04/2024	that month
--date 28/04/2024	that day
--date 2023..2024	both years
--date 04/2024..06/2024	April to June
--date 2023..	2023 onwards
--date ..2020	everything up to the end of 2020
--date 30d	the last 30 days; also 2y, 1w, 1m

Both ends of a range may be a year, a month or a day, and they need not match. An upper bound always includes the whole period it names, so ..2020 runs to 31 December.

04/28/2026 is rejected rather than guessed at — 28 is not a month, and the error says so. So is anything else it cannot read: 2024-o6-01 with a letter for a zero is an error, not a silent answer about January. A range that runs backwards (2024..2020) says so rather than returning nothing.

In browse, d opens the same thing as a prompt: type 2023..2024, enter to apply, an empty prompt to clear, esc to cancel. A misparse keeps the prompt open with the error in the footer, so you correct it rather than retype it. The active range shows in the header — including one that came from --date on the command line, which was previously invisible once you were inside.

Search scope

By default a query matches title, authors and abstract. -t / --title narrows it to title and authors only — useful when searching for a person, where abstract matches are mostly papers citing them rather than papers by them.

eprint Boudgoust          # 22 hits, four of them papers *citing* her
eprint Boudgoust -t       # 18 — title and authors only

In browse, press t to toggle scope and watch the result count change live.

Scripting

--json emits structured results; colour and hyperlinks are suppressed automatically when output is piped.

eprint "threshold signature" --json | jq -r '.[].url'
eprint zk --color | less -R        # force colour through a pager

How the index stays current

Searching checks the index age. If it is more than 24 hours old, a background refresh is spawned and your search returns immediately against existing data — it never blocks. An incremental update takes a second or two. Use --no-update to suppress it, or eprint update to refresh on demand.

An incremental harvest asks for changes since the earlier of two things: the last harvest, and the newest paper actually in your index — less a two-day overlap. Both halves matter. The harvest clock alone cannot be trusted, because a refresh that finds nothing still advances it, and once it runs ahead of the data every later request starts after the papers you are missing; that cost this author's own index a week of papers while status cheerfully reported a fresh harvest. Bounding the window by the newest paper you hold makes the request describe your data instead, so a gap of any size closes itself on the next update. Withdrawn papers arrive as OAI-PMH tombstones and are deleted.

With a background updater installed the index is refreshed every 30 minutes regardless, so the staleness check above never fires. It is the OS's own scheduler, so there is nothing to start and nothing left running, and it catches up after the lid has been shut.

Scope and licensing

Metadata comes from the archive's OAI-PMH endpoint, which the site publishes for exactly this purpose. Search covers title, authors, abstract and IACR category. Author-supplied keywords are not part of the oai_dc feed and so are not indexed; abstracts generally contain the same terminology.

Full-text PDFs are licensed individually per paper and are never fetched by this tool. eprint open hands the URL to your browser, so the download happens in a normal browser session under that paper's licence; the tool only files a copy your browser has already saved. eprint show displays a paper's licence (CC-BY-4.0, CC-BY-NC-ND-4.0, CC0, …) so you can see the terms before opening.

The harvester identifies itself honestly by User-Agent, paces requests, and honours Retry-After on 503/429.

Storage

Everything lives in two places, both safe to delete — the index rebuilds from scratch.

What	macOS	Linux
Index + citation keys	~/Library/Application Support/eprint/eprint.db	$XDG_DATA_HOME/eprint/eprint.db
Config	~/.config/eprint/config.toml	~/.config/eprint/config.toml

--notify and --launcher add files outside both, each removed by the same flag set to off:

What	macOS	Linux
Background updater	~/Library/LaunchAgents/local.eprint.update.plist	~/.config/systemd/user/eprint-update.{service,timer}
Updater's error log	~/Library/Logs/eprint-update.log	journalctl --user -u eprint-update
Notification sender	~/Library/Application Support/eprint/eprint.app	— (notify-send is used directly)
Launcher	~/Applications/eprint.app	~/.local/share/applications/eprint.desktop

The database is roughly 94 MB with metadata only, or ~112 MB once CryptoBib entries are stored. Saved PDFs live separately, in ~/Documents/eprint/. Override the locations with $EPRINT_DB and $EPRINT_CONFIG.

Development

cargo build --locked            # debug build
cargo build --release --locked  # optimised build
cargo clippy --all-targets      # lints
cargo test --locked             # unit tests
cargo fmt                       # formatting

GitHub Actions runs the build and tests on macOS and Ubuntu, and clippy with warnings denied on Rust 1.86 — the version the lockfile is held back for. Ubuntu is the half that cannot be exercised by hand on a Mac, which is the point.

The launcher's icon is a committed asset, embedded with include_bytes!; regenerate it by hand with swift assets/icon.swift assets/eprint.icns, which is why nothing in the build needs Swift.

Pass EPRINT_DB=/tmp/scratch.db to work against a throwaway index instead of your real one. eprint config --notify honours it too, baking whatever EPRINT_* overrides are set into the scheduled job, so testing this cannot point a background harvest at your real index.

Layout

The crate is a library plus the binary that uses it. The division is not decorative: everything in src/lib.rs is free of the terminal — no ratatui, no ANSI, no std::process, no current_exe — so the index, the harvester and the author-name table can be reused without dragging a terminal along, while every decision about a terminal stays with the binary.

The library — the index, and what is true about it:

File	Responsibility
src/lib.rs	The library root, and why the boundary is where it is
src/db.rs	SQLite schema, migrations, FTS5 query building, the watch cache
src/harvest.rs	OAI-PMH harvester, the incremental window, progress and cancellation
src/feed.rs	What a bare eprint shows: the batch, the top-up, the replay, the label
src/bib.rs	CryptoBib fetch, BibTeX parsing, @String macro expansion
src/names.rs	Who an author is: folding, the name table, the one matching predicate
src/dates.rs	Civil-date arithmetic, ISO storage, the day-first grammar
src/config.rs	Config file reading and writing (settings and watches)
src/text.rs	Layout with no opinion about colour: wrapping, bylines, dates, --json

The binary — the terminal, and the integrations it installs:

File	Responsibility
src/main.rs	CLI definition, command dispatch, config/flag resolution
src/render.rs	Inline output: styling, highlighting, hyperlinks
src/tui.rs	Interactive browser (ratatui)
src/theme.rs	Colour palettes shared by both front-ends
src/pdf.rs	The local PDF library: filing and finding saved papers
src/completions.rs	The zsh and bash functions, and installing them
src/notify.rs	Desktop banners: which tool posts them, and what they say
src/desktop.rs	The launchd/systemd/.app/.desktop files, and installing them
assets/icon.swift	Generates assets/eprint.icns; run by hand, not by cargo

Two independent data sources feed one SQLite database: ePrint's OAI-PMH endpoint supplies paper metadata (papers, plus an FTS5 index), and CryptoBib supplies citation keys (bib). Neither is bundled with the tool; both are fetched at runtime into your own database. Schema changes are applied by migrate() in src/db.rs, which runs on every open, so existing databases upgrade in place.

Licence

MIT — see LICENSE.

The two data sources have their own terms, independent of this tool:

• IACR Cryptology ePrint Archive — papers are licensed individually; each result shows its licence. Metadata is taken from the OAI-PMH interface the archive publishes for that purpose.
• CryptoBib — fetched at runtime and cached locally. Nothing from it is redistributed with this tool.