An information-theory Wordle solver for Windows, macOS, and Linux. A pure-Rust engine (entropy + frequency/recency-weighted scoring) behind a Tauri 2 desktop shell with a React + TypeScript UI.
It implements the Shannon-entropy method from Aladaileh et al. (2026), Solving Wordle Using Information Theory, plus improvements beyond the paper: frequency-weighted priors, recency-weighted past answers, an expected-guesses objective, exploratory guessing, and an optional two-step lookahead. On the original answer set it reaches a 100% solve rate at ~3.47 average guesses (the exact dynamic-programming optimum is ~3.42).
- Assist mode: play Wordle anywhere (the real NYT puzzle, an archive, a clone, a Discord bot) and mirror the results into the tile board; Wordtropy suggests the next best guess. You can type a guess or click a suggestion to drop it on the board, set each tile's color, and confirm (you can fix a color in the confirmation).
- Practice mode: play a hidden word right in the app; it colors your guesses automatically, with the same hints available.
- Suggestions ranked by the engine, each showing expected information (bits), an estimate of
guesses-to-go, and a badge:
- green target = a possible answer (could win outright)
- blue funnel = an information probe (likely not the answer, but best splits the remaining words to eliminate/confirm many at once)
- amber clock = a past NYT answer (with how long ago; down-weighted, never today's word)
- Candidates: every still-possible word with its likelihood, frequency, and expected information; searchable, sortable, and can hide past answers.
- Answer pools: Original (2,315 pre-NYT answers; reproduces the paper), Curated (original plus every word NYT has used), or Full list (any valid guess; the default). All pools auto-include NYT's latest answers, and the solver auto-widens to the full word list if a pool ever can't contain the answer, so it never dead-ends.
- Recency data and the cheat stay current via a launch-time fetch from the official NYT Wordle API (offline-safe).
Each guess partitions the remaining answers into buckets by the color pattern it would produce; a guess's entropy is the expected bits of information it reveals. Picking the highest-entropy word is the paper's method. Beyond that, Wordtropy adds:
- Frequency-weighted priors: candidates are weighted by word frequency (Google Books + OpenSubtitles), so common words are treated as more likely answers.
- Recency weighting: recently-used NYT answers are down-weighted with an age decay (today's answer is never penalized).
- Expected-guesses objective: instead of raw entropy, the solver minimizes estimated total guesses and prefers a word that could win now.
- Exploratory guessing: when many candidates remain, a non-candidate word is allowed if it splits the set better.
- Two-step lookahead (optional): refines the shortlist with depth-2 search.
A precomputed pattern[guess][answer] matrix (built with rayon, cached to app-data) accelerates
scoring, and the opener is cached for instant new games.
Download an installer from the Releases page and run it:
- Windows:
Wordtropy_x64-setup.exe(NSIS) orWordtropy_x64_en-US.msi - macOS: the universal
.dmg - Linux:
.AppImage(most portable) or.deb/.rpm
Tested on Windows only. The macOS and Linux builds are produced by CI and have not been tested by the author, so they may have rough edges (the custom title bar in particular is not yet native on macOS). Bug reports and feedback are welcome.
The apps are currently unsigned: Windows shows a SmartScreen prompt (More info, then Run
anyway), macOS needs right-click then Open the first time, and the Linux .deb/.rpm require
WebKitGTK.
Prerequisites: Rust, Node 20+, and pnpm. On Linux you also need the Tauri system dependencies (WebKitGTK and friends; see the Tauri docs).
git clone https://github.com/BrokenHypocrite/wordtropy.git
cd wordtropy
pnpm install
pnpm tauri dev # run the app in development
pnpm tauri build # build installers for the current OSThe engine and benchmark are pure Rust and run without the UI:
cargo test --workspace # all tests
cargo run -p wordtropy-bench --release -- --gate # weighted vs greedy acceptance gate
cargo run -p wordtropy-bench --release -- --pool answers --mode weightedTauri cannot cross-compile (it links each OS's native webview), so each platform builds on its
own machine. The included GitHub Actions workflow (.github/workflows/build.yml) does this for
you: push a tag (git tag v0.1.0 && git push --tags) to build Windows, macOS (universal), and
Linux and create a draft GitHub Release with the installers, or run it manually to get build
artifacts.
engine/: pure-Rust solver library (wordtropy-engine). Word packing, pattern computation, entropy/scoring, priors, recency, pattern matrix, session state. Unit-tested and benchmarkable without Tauri.bench-harness/: headless benchmark binary (solve rate, average guesses, histogram).src-tauri/: the Tauri 2 desktop app. IPC commands, app state, NYT-API refresh.src/: the React + TypeScript UI (Vite).
- Aladaileh, Stephens, Alqaisi, Wu (2026). Solving Wordle Using Information Theory, NEJCS: https://orb.binghamton.edu/nejcs/vol8/iss1/6
- 3Blue1Brown, Solving Wordle using information theory: https://www.youtube.com/watch?v=v68zYyaEmEA
- Bertsimas & Paskov (2022). An exact and interpretable solution to Wordle (the ~3.42 optimum).
- Pattern-computation reference,
jonhoo/roget: https://github.com/jonhoo/roget - Word lists (original Wordle), cfreshman gist: https://gist.github.com/cfreshman/a03ef2cba789d8cf00c08f767e0fad7b
- Daily answer history, the official NYT Wordle API: https://www.nytimes.com/games/wordle/index.html
- Frequencies: Google Books Ngram (via 3Blue1Brown) and OpenSubtitles
(
hermitdave/FrequencyWords): https://github.com/hermitdave/FrequencyWords
Data provenance and licensing for the bundled word lists, frequencies, and answer history are
documented in engine/assets/SOURCES.md.
The Wordtropy source code is released under the MIT License.
Bundled data assets keep their own licenses (see
engine/assets/SOURCES.md):
engine/assets/freq_subs.tsv(OpenSubtitles frequencies) is CC-BY-SA-4.0.engine/assets/freq_google.tsvcomes from 3Blue1Brown's data (MIT) and Google Books Ngram.- The Wordle word lists and the NYT answer history are factual game data, widely mirrored.
Wordle is a game by The New York Times. This project is not affiliated with, endorsed by, or sponsored by NYT.