Pronounce / pronounce-mcp

🚀 Try it in 30 seconds

git clone https://github.com/anzy-renlab-ai/pronounce.git
cd pronounce && ./install.sh
say-it kubectl

That's it. Now try say-it GIF, say-it nginx, say-it Pydantic, say-it --why JSON, or say-it quiz for a 10-question challenge. Linux users: install espeak-ng (sudo apt install espeak-ng) and the CLI just works. Windows: same CLI under WSL or git-bash + PowerShell. Or skip install and use the browser at pronounce.renlab.ai.

⭐ If say-it kubectl saves you one cringey standup moment — star the repo. It nudges more devs to contribute their favorite mispronounced project name.

Related MCP server: wordorb

🏆 The developer pronunciation scoreboard

1,880 entries — 1,260 carry a citable source — 101 settled by the creator themselves, 174 the community still argues about. The famous ones:

✅ Settled — the creator said so

⚔️ Still contested — both readings are in active use

Every cell has IPA and audio; 1,260 also carry a citable source. Hear all 1,880 →

Disagree with one? That's the whole point — open a PR with your reading and a source. The argument is the dataset.

What you're actually getting

• 1,880 entries — 1,260 carry a citable source. Confidence-tagged (creator-clarified / community-consensus / contested), each with a citable URL where one exists (we'd rather leave it blank than fabricate one). Wilhite said GIF is "jif" at the 2013 Webby Awards. Crockford says JSON is "JAY-son" (RailsConf 2009). RFC 7519 says JWT is "jot". The dictionary cites them.

• Multi-reading audio. For words where the debate is real — GIF, SQL, GUI, char, regex — the CLI chains the alternates after the primary with a spoken "or:" so you hear the debate without staring at the terminal. --solo skips the tail once you've internalized it.

• One Bash CLI, no npm runtime. No sudo, no framework bootstrap, no surprises. It detects macOS say, Linux espeak-ng/espeak, or Windows PowerShell System.Speech. The repo also ships a Claude Code skill and an MCP server so your AI answers "how do you pronounce X?" with audio, not a phonetic guess.

• Canonical website audio. The site plays each committed canonical MP3 first and uses Web Speech only as a fallback if playback fails.

$ say-it --why JSON
word              JSON
ipa               /ˈdʒeɪsən/
respelling_us     jay son
confidence        contested
source            Wikipedia § Pronunciation
url               https://en.wikipedia.org/wiki/JSON#Pronunciation

Why not just Google?

Because Google gives you 47 Reddit arguments and a YouTube clip you have to unmute. IPA gives you /ˈkuːb kənˌtroʊl/ — a reference, not a teacher.

You don't need a phonetic transcription. You need to hear the word. Twice. Maybe three times. Done.

say-it ships a community-maintained dictionary of how engineers actually say the names that trip everyone up — and feeds the intended respelling to your OS's text-to-speech engine, so kubectl comes out as koob-control, not whatever your computer guessed from the letters.

Famous moments

Some pronunciations aren't opinions — the creators settled them. The dictionary cites every one:

Run say-it --why <word> to see the citation when an entry has a source_url; unsourced entries are left blank rather than given a fabricated URL.

Install

brew install anzy-renlab-ai/tap/say-it   # Homebrew on macOS

Or the "Try it in 30 seconds" block above. ./install.sh drops:

• the CLI at ~/.local/bin/say-it,

• the pronunciation dictionary at ~/.local/share/say-it/pronunciations.tsv,

• if you use Claude Code, a pronounce-word skill at ~/.claude/skills/pronounce-word/ so any "how do you say X?" prompt to your AI gets answered with audio instead of IPA,

• the same skill into ~/.agents/skills/ (Codex CLI) and ~/.kiro/skills/ (Kiro) when those dirs exist — it's the cross-tool Agent Skills standard.

Make sure ~/.local/bin is on your $PATH. Linux also works — install espeak-ng (sudo apt install espeak-ng / brew install espeak-ng). Windows: WSL or git-bash + PowerShell. Or skip install entirely with the browser version at pronounce.renlab.ai.

Usage

say-it kubectl                    # primary × 3, then "or: <alt>" for each alternate
say-it --solo kubectl             # primary only — silence the "or:" tail
say-it --alt GIF                  # focus on the first alternate
say-it --alt 2 GUI                # focus on the Nth alternate (1-indexed)
say-it --all SQL                  # primary AND every alternate, each repeated
say-it --no-dict kubectl          # bypass the dictionary entirely
say-it --why JSON                 # show IPA, source URL, category, confidence
say-it list                       # every word in the dictionary
say-it search redis               # grep the dictionary (case-insensitive)

say-it -n 5 Pydantic              # 5 repetitions instead of 3
say-it -r 110 Knative             # slower (110 wpm; default is 130)
say-it -o /tmp/word.aiff Postgres # save to file instead of playing
say-it --list                     # all macOS voices

The default macOS voice is Samantha (General American); -v <voice> selects another macOS voice. Linux and Windows use their detected backend's English voice. The dictionary is GenAm-only, by design.

Claude Code integration

You:    kubectl 怎么读？
Claude: 🔊 (plays "koob-control" three times)
        /ˈkuːb kənˌtroʊl/ — "KOOB-control". Kelsey Hightower says it
        this way (KubeCon talk). "Cube-cuddle" is an alternate —
        try `say-it --alt kubectl` to hear it.

Once installed, the pronounce-word skill auto-triggers on:

• X 怎么读 / X 怎么发音 / 读一下 X

• how do you pronounce X / pronounce X / how do you say X

Your AI replies with sound, not just a phonetic guess. Skill file: skills/pronounce-word/SKILL.md.

Not on Claude? Same skill drops into Codex CLI (codex plugin marketplace add anzy-renlab-ai/pronounce) and Kiro (~/.kiro/skills/), and the MCP server covers Claude Desktop, Cursor, Continue, Zed, Cline & friends — full matrix in integrations/.

VS Code extension

Hover over any tech word in any file — see the IPA, hear the pronunciation. Same 1,880-entry dictionary as the CLI, JSON-bundled at build (zero runtime parse cost).

# Cursor / VSCodium / Zed / Gitpod / Theia / code-server (Open VSX)
code --install-extension sayit.pronounce
# or the listing: https://open-vsx.org/extension/sayit/pronounce

Now live on Microsoft Marketplace too: ext install sayit.pronounce in VS Code. See marketplace listing.

• Hover over kubectl, YAML, Ghostty, wagmi… → tooltip with IPA + 🔊 Play + ★ Star link.

• ⌘⇧' — speak selection.

• Status bar 🔊 sayit — click to speak the current selection.

• Welcome walkthrough — 4-step onboarding on first install.

• Pronounce: Search dictionary… — fuzzy-find all 1,880 entries.

Source: integrations/vscode/. Cross-platform as of v0.3 — macOS say, Linux espeak-ng, Windows PowerShell.

Chrome / Edge / Brave extension

Click any tech word on any webpage → popup with IPA + audio. Same 1,880-entry dictionary. The extension speaks through Web Speech; pronounce.renlab.ai plays the committed MP3 corpus first and reserves Web Speech for fallback. Sideload only for now (not yet on Chrome Web Store).

Download pronounce-chrome-0.3.1.zip → unzip → chrome://extensions/ → Developer mode → Load unpacked.

Source: integrations/chrome/.

How the dictionary works

data/pronunciations.tsv is the single source of truth — tab-separated, 1,880 entries, covering:

• Cloud / DevOps: kubectl, nginx, Kubernetes, helm, Istio, Envoy, Prometheus, Grafana, Terraform, Argo, Knative, etcd, containerd, runc, Podman, ...

• Languages / Frameworks: Django, Vue, Vite, Pydantic, Bun, Deno, Hugo, Hono, Caddy, Svelte, Astro, Pinia, ...

• Databases: PostgreSQL, Postgres, SQLite, MySQL, MongoDB, Cassandra, Redis, Ceph, ScyllaDB, ClickHouse, DuckDB, ...

• CS jargon / acronyms: GIF, JSON, SQL, GUI, GNU, char, regex, sudo, tmux, chmod, WYSIWYG, ASCII, enum, NaN, SaaS, PaaS, ...

• Distros / tools: Linux, Debian, Ubuntu, Arch, Nix, LaTeX, TeX, emacs, zsh, ...

Each entry has 10 columns: word | ipa | respelling_us | alt_ipa | alt_respelling_us | source_url | source_label | category | confidence | notes. respelling_us is plain English-like input passed to the detected OS TTS backend, so the intended reading is spoken instead of asking the engine to guess from the project name.

Local override: drop a ~/.config/say-it/pronunciations.local.tsv and it takes precedence.

What works today

• ✅ macOS — built-in say.

• ✅ Linux — espeak-ng (preferred) or espeak.

• ✅ Windows — PowerShell System.Speech from the same Bash CLI under Git Bash/MSYS2/Cygwin.

• ✅ 1,260 of 1,880 entries carry a citable source — the rest are confidence-tagged, no fabricated citations.

• ✅ Audible multi-reading awareness — contested words audibly chain alternates with "or:".

• ✅ --alt [N], --all, --solo, --why, --json, --md, --no-dict, list, search, quiz, repl, stream, doctor, export, benchmark, badge, cheatsheet.

• ✅ Claude Code skill + MCP server for AI-side pronunciation questions.

• ✅ Browser PWA — installable, offline-capable, instant search, voice-mic search, interactive quiz.

• ✅ Editor integrations — Raycast, Alfred, VS Code, Cursor, Codex, Kiro, Continue.

• ✅ 🌐 Live site — pronounce.renlab.ai (1,880 words browsable with audio; citations on the 1,260 sourced entries) + /zh (Chinese landing).

What's coming

See DESIGN.md for the architecture.

• ☁️ Cloud TTS (opt-in ElevenLabs / OpenAI) for the names native TTS still mangles.

• 📚 Anki export for vocabulary drills.

• 🌗 Light theme on the v2 homepage (already shipped on word/SEO pages).

Contributing

Two things we want most:

1. Pronunciation entries. Open a PR adding a row to data/pronunciations.tsv. Required columns: word, ipa, respelling_us. Highly preferred: source_url (creator interview, conf talk, official FAQ — anything verifiable). Contested readings are welcome; put the rival in alt_* columns and we'll wire --alt through.

2. Backend quality. macOS (say) is the gold standard; the Linux (espeak-ng) and Windows (PowerShell) backends ship but are best-effort — help us close the gap. See DESIGN.md §Backends.

Keep it tiny. Keep it dep-free where possible. Keep the defaults opinionated (3 reps, GenAm, Samantha voice).

⭐ Support — start with a star

The dictionary is free and MIT. The single highest-leverage thing you can do is star the repo — it's the signal that pulls in more contributors, more PRs, more creator-clarified entries.

• ⭐ Star on GitHub → — one click, no signup, biggest effect.

Optional, if it's saved you real standup pain:

• ☕ Coffee on Ko-fi — pays for one new entry per cup.

• 💚 Sponsor on GitHub — recurring tier (pending Sponsors approval).

Dollars cover hosting (Vercel/Cloudflare/Open VSX), domain renewals, MiniMax narration credits for promo videos, and time to track down creator citations for new entries.

Contributors

Every entry, source upgrade, and skill fix counts. Open a PR — your face shows up here.

License

MIT — see LICENSE.

IPA is a reference. Audio is a teacher.