Harness Discussion

Select text in any Windows or Mac app and hear it read aloud with the voice from your Cursor or Codex sign-in.

Harness Discussion reads with Cursor's voices through your Cursor desktop sign-in, or with the Codex realtime voice from your ChatGPT sign-in through codex app-server. No API key. Local voices take over when neither can speak.

Unofficial. Not affiliated with Cursor or OpenAI.

Download for Windows · Download for Mac · Website · Report a bug

What it does

• Works in most apps. Drag or double-click to select text in a browser, Notepad, VS Code, or most other apps. A rounded bar appears next to the selection without taking focus: Play, Summarize, voice, speed, …, and Close. Language and Settings are in ….
• Reads with the sign-in you already have. First Cursor, with any of its 28 voices, through the Cursor desktop app's sign-in. Then the Codex realtime voice through the Codex CLI sign-in, which uses your ChatGPT plan's Codex access and counts toward its limits.
• Then Kokoro, on your computer. Kokoro is an open-weight neural voice that runs offline. Settings → Voices downloads it once (about 120 MB). It reads English, Spanish, and French.
• Then local voices. If none of those can speak, Edge neural voices take over, then the offline Windows voices. On a Mac, the built-in macOS voice takes over.
• You set the order. Settings → Voices lists the sources in reading order. Turn each one on or off, move it up or down, pick its voice, and test it. A source that needs installing or a sign-in shows a button that opens the download page or starts the sign-in. The bar's source menu moves a source to the top. Picked during a read, it takes over from the sentence being read.
• Summarize. ✦ next to Play summarizes the selection on your computer with Qwen3-4B through llama.cpp, then reads the summary. The first click asks to download the model once (2.5 GB, verified by SHA-256). Settings → Summarize shows its status and removes it.
• Check as I type. On with Writing: when you pause, the default mode's changes to the line are underlined. Click one to apply it, or press the rewrite shortcut with nothing selected to apply all.
• Rewrite in place. Over text you can edit, ✎ on the bar, Ctrl+Shift+E on Windows, or ⌥⌘E on a Mac rewrites the selection with the default mode and pastes it back. Modes are Fix errors, Message, Email, Improve writing, Clean dictation, and your own. A toast offers Undo, Review (keep or revert each change), and another mode. Click × to dismiss the result on either platform. Press the rewrite shortcut twice quickly to pick a mode first. Cancel on the toast, or Esc, stops it. Cursor, Codex, then the local Qwen3-4B model rewrite, in the order you set. On for new installs; turn it off in Settings → Writing.
• About you. A short note about you, plus today's date and time zone, goes with every rewrite and summary. Set it in the tour or Settings → General.
• Dictation. Hold Ctrl+Shift+D on Windows (⌥⌘D on Mac), speak, then release. Local (Whisper Small, downloaded in Settings → Dictation) is tried first, then enabled cloud sources; the dictation mode tidies the result before insertion, preferring the enabled, downloaded Local writing model. Pick the input device in Settings → Dictation; input bars show while recording. Esc or × stops it. Clean dictation removes fillers and formats spoken steps as numbered lists, including corrected numbers. Settings → Dictation sets its own mode; custom prompts come from Writing → Modes. Turn off Tidy dictation after transcribing for raw text. A short tap does nothing. Uses the Writing switch. Windows support is experimental; live English, Hebrew, and mixed speech acceptance is pending (#92).
• Pause, resume, stop, skip. ⏸ pauses and ▶ continues from the same spot. ■ stops. ⏮ and ⏭ read again from the previous or next sentence.
• Read shortcut. Ctrl+Alt+R on Windows, ⌥⌘R on a Mac, reads the selection at once and shows the bar at the pointer. Press it again to stop. Change it or turn it off in Settings → Reading.
• Speed on the bar. 0.5× to 2×, saved and shared with Settings.
• Read-along captions. The current sentence appears above the bar as it plays.
• Six languages. Hebrew, English, Arabic, Russian, Spanish, and French.
• Updates itself if you want. Optional automatic updates from GitHub releases, verified by SHA-256.
• Speech-to-text server (optional). An OpenAI-compatible local endpoint that transcribes with the same Codex sign-in. See STT.md.

Images show the current Windows UI with sample text. How the promotional images were made.

Install

1. Download HarnessDiscussion-Setup.exe and run it. It installs for your user only and needs no admin rights.
2. The installer isn't code-signed, so SmartScreen may warn you. Choose More info, then Run anyway.
3. Optional: sign in to the Cursor desktop app, or install the Codex CLI (npm i -g @openai/codex) and sign in once with codex login. Without either, the app uses the Edge and Windows voices.
4. On first launch, a short tour has you try reading, rewriting, Review, dictation, and Check as I type on sample text, change the shortcuts, and add an About you note. Take the Tour in the tray menu replays it.

Requirements: Windows 11, 64-bit.

macOS

1. Download HarnessDiscussion.dmg, open it, and drag Harness Discussion to Applications.
2. Open Harness Discussion from Applications. If macOS says it "could not verify" the app, the build isn't notarized yet:
   • On macOS 15 and later, choose Done, open System Settings > Privacy & Security, scroll to Security, choose Open Anyway next to Harness Discussion, then confirm with your password.
   • On macOS 13 and 14, right-click the app, choose Open, then Open again.
   • Or run xattr -dr com.apple.quarantine "/Applications/Harness Discussion.app" in Terminal, then open the app.
3. Harness Discussion lives in the menu bar. The tour's first step asks for Accessibility access, which the app needs to read the selected text. A guide window opens System Settings for you and closes once access is granted. While access is missing, the menu bar icon shows a warning.
4. Optional: sign in to Cursor, or install the Codex CLI and run codex login, as on Windows.

Requirements: macOS 13 or later on Apple silicon.

Edge voices are Windows only for now. Start Harness Discussion when I log in adds a LaunchAgent. Settings are kept in ~/Library/Application Support/harness-discussion. Updates are not installed automatically: Settings shows when a new version is out, and Download downloads it, checks its SHA-256 digest, replaces the app, and reopens it. If the app can't replace itself (it runs from the dmg or a folder you can't write to), Download opens the release page.

After an update, macOS can stop trusting the app while System Settings still lists it as allowed. The app detects this, and the guide's Fix permission button clears the old entry so you can grant it again.

Upgrading from Codex Speak or Select to TTS? It's the same app, renamed. On Windows, the update or new installer replaces the old app and keeps your settings, sign-in startup choice, and speech-to-text URL. On a Mac, open the new app once: it quits Codex Speak, moves its settings, and replaces its login item. macOS asks for Accessibility again because the app's identity changed. Then delete Codex Speak from Applications.

Use

1. Select text with the mouse (drag or double-click).
2. Press ▶ on the bar. The language menu defaults to Auto.
3. Pick a speed from the bar's speed menu, and a source from its source menu.
4. Selecting new text while paused replaces the paused read.
5. Or press the read shortcut (Ctrl+Alt+R, or ⌥⌘R on a Mac) to read the selection at once.
6. Over text you can edit, press ✎ or the rewrite shortcut (Ctrl+Shift+E on Windows, ⌥⌘E on Mac). Hold the dictation shortcut (Ctrl+Shift+D on Windows, ⌥⌘D on Mac) to dictate.

On a Mac, the menu bar icon's menu shows what is playing and has Pause, Stop, Speed, and Voice Source. Show Popup on Selection turns the bar off until you turn it back on or restart the app.

Open Settings from the bar's … menu, the tray or menu bar icon, or by starting the app again from the Start menu.

Where your text goes

Voice	Needs	Speed	Where your text goes
Cursor	Cursor desktop app, signed in	0.5× to 2×	Cursor, through your Cursor sign-in
Codex realtime	Codex CLI signed in with ChatGPT	0.5× to 1×. Above 1×, the next source reads	OpenAI, through your Codex sign-in
Kokoro	A one-time 120 MB download from Settings	0.5× to 2×	Stays on your computer
Edge neural (Windows)	Internet	0.5× to 2×	Microsoft's Edge read-aloud service
Windows	Nothing, it's built in	0.5× to 2×	Stays on your PC
macOS	Nothing, it's built in	0.5× to 2×	Stays on your Mac

Rewrite and dictation	Needs	Where your text or audio goes
Cursor	Cursor CLI (cursor-agent), signed in	Cursor, through your Cursor sign-in
Codex	Codex CLI signed in with ChatGPT	OpenAI, through your Codex sign-in
Local (rewrite only)	The Summarize model	Stays on your computer

• For local read-aloud, enable only Kokoro and Windows or macOS in Settings → Voices. Writing and Dictation have separate source settings.
• Password fields are never read or rewritten.
• For local rewrites, turn off Cursor and Codex in Settings → Writing and use Local. Dictation audio goes to the first ready source (Local by default) and is not saved. For offline dictation and Tidy, download both local models and turn off Cursor and Codex under both Writing and Dictation. Enabled cloud fallbacks can receive audio or the transcript if a local stage fails.
• A rewrite is pasted only if the same app, window, and selection are still there. Your clipboard is put back afterwards.
• On Windows, the app reads accessible selections without automatic clipboard copying. Apps that hide their selection cannot be read, and list numbers may be omitted.
• The log records the engine, timing, and errors, never the text itself.
• Kokoro has no Hebrew, Arabic, or Russian voice, so the next source reads those.
• In Auto, Cursor and Codex work out the language themselves. Kokoro, Edge, and Windows pick Hebrew, Arabic, or Russian voices from the script and use English for everything else. Choose Spanish or French from the menu for those voices.

Settings

• Start when I sign in to Windows: adds or removes a Task Scheduler logon task named "Harness Discussion".
• Check and install update / Install updates automatically: see Updates.
• Speed: 0.5× to 2×. Changes apply to the next read.
• Sources: under Voices, the voice sources in reading order, by default Cursor, Codex, Kokoro, Edge, then Windows (Cursor, Codex, Kokoro, then macOS on a Mac). Kokoro is skipped until you press its Download…. The model goes to %LOCALAPPDATA%\harness-discussion\kokoro (~/Library/Application Support/harness-discussion/kokoro on a Mac), and its size and SHA-256 digest are checked. Each row shows whether the source is ready, its voice where it has one, an on/off box, Test, and ↑ ↓ to reorder. If one source can't read, the next one does.
• Captions: font size (8 to 24 points), optional background, and background opacity. Changes apply to the visible caption right away.
• Read apps that hide their selection: on Windows, unavailable to protect your clipboard. The app reads selections exposed through accessibility. On a Mac, this option copies with ⌘C and then puts your clipboard back.
• Prepare speech as soon as I select text: off by default. When Cursor or Edge reads first, the app requests the opening's audio as soon as the bar appears. This sends those sentences to Cursor or Microsoft before you press play.
• Speech-to-text server: see STT.md.
• Writing: turns rewriting on or off, and sets the default mode, the rewrite shortcut, and the sources in order with each one's model and effort. The modes editor adds, renames, and deletes your own modes. Built-in modes keep their name, and Reset puts back their instructions.
• Dictation: source order (Local first by default), local model download, cleanup and local preference, its own mode (including custom modes), hold shortcut, and input device. Input bars show while recording; × cancels. Dictation inserts quietly or copies to the clipboard without a safe target. Windows Microphone settings opens system privacy. If microphone shutdown stalls, captured audio still transcribes. Restart if the microphone stays unavailable.
• Learned words: on by default in Dictation. For 30 seconds after a confirmed paste, useful spelling corrections in that field can join your vocabulary. A notification offers Undo. Turn learning off, edit individual lines, or clear the list in Settings → Dictation. Saved words help later Local/Codex transcription and dictation tidy; cloud sources can receive them. .

Settings are saved in %APPDATA%\harness-discussion\settings.json, next to a small harness-discussion.log.

Cursor, Edge, Windows, and macOS voices support the full range. Codex streams speech in real time, so it can only be slowed down. Above 1×, the next source in the order reads. With only Codex turned on, it reads at its natural pace.

Check and install update checks the latest stable GitHub release, verifies the installer's size and SHA-256 digest, installs it into the current app folder, and relaunches. It waits for speech, including paused speech, and active transcriptions to finish.

Install updates automatically is off by default. When on, the app checks shortly after startup and every 24 hours, and installs when idle. Turning it off cancels a pending automatic installation. Settings and your sign-in startup choice are kept.

If a check or download fails, the app keeps running and Settings shows the error. Downloaded installers and setup.log are kept under %LOCALAPPDATA%\harness-discussion\updates for diagnosis and removed on uninstall. The digest checks that the download matches GitHub's release metadata. The installer is unsigned, so the updater trusts this repository's release account.

Update controls are disabled when running from source.

Use Settings > Apps > Harness Discussion. This removes the app, its sign-in task, the Kokoro model, and its caches in %TEMP%. Your settings and log in %APPDATA%\harness-discussion are kept.

Known limitations

See docs/known-limitations.md: Codex and Cursor speech can break with their updates, how long selections are split, and a few platform quirks.

Run from source

uv sync
uv run harness-discussion     # or start .venv\Scripts\harness-discussion.exe (no console window)
uv run python -m unittest discover -s tests

CONTRIBUTING.md covers the dev setup, building the installer, releases, and a map of the code.

Code signing policy

Windows code signing is pending. The project is applying for free code signing from SignPath Foundation, and current installers are still unsigned. When signing starts, Windows releases will say:

Free code signing provided by SignPath.io, certificate by SignPath Foundation.

• Authors and committers: tzachbon
• Reviewers: tzachbon
• Approvers: tzachbon

Only the app's own executable and its installer will be signed, from GitHub Actions builds of this repository, and every release needs a manual approval. See the privacy policy for where your text goes.

Contributing

Bug reports, fixes, and new language voices are welcome. Start with CONTRIBUTING.md. Report security issues privately as described in SECURITY.md.

License

MIT. Not affiliated with Cursor, OpenAI, Microsoft, or Apple.

The Kokoro voice uses kokoro-onnx (MIT) and the Kokoro-82M weights (Apache-2.0). The installers bundle espeak-ng for pronunciation, which is licensed under GPL-3.0.