Talk, point, paste.
Show an AI what you mean instead of describing it.
or paste one line: see Install
Press one shortcut, say what you see, and point at or mark up the screen. When you stop, Pointcast pastes one prompt, your words with the annotated screenshots where you took them, into the app you were using: ChatGPT, Claude, Cursor, Claude Code, Codex, Slack or any text box.
It is a small native app for the macOS menu bar and the Windows tray. Speech is transcribed on your computer. Nothing is uploaded.
The fastest way: copy one line, paste it, press Return. It downloads the latest version, checks it, installs it and opens it. No warnings, no files to open.
macOS 12+: open Terminal (Spotlight: "Terminal") and paste:
curl -fsSL https://raw.githubusercontent.com/naidx0/pointcast/main/install.sh | bashWindows 10 or 11: open PowerShell (Start: "PowerShell") and paste (no admin rights needed):
irm https://raw.githubusercontent.com/naidx0/pointcast/main/install.ps1 | iexOr use the Download buttons above (every version is on Releases). Pointcast is not signed with a paid Apple or Microsoft certificate, so a downloaded installer gets a warning the one-line install avoids: on macOS, drag Pointcast to Applications, then right-click it and choose Open the first time (on macOS 15: System Settings → Privacy & Security → Open Anyway); on Windows, choose More info → Run anyway.
A setup assistant then takes about a minute: allow access, pick a speech model, try the shortcuts and make a practice recording. Pointcast opens at login from then on and updates itself from Settings → General.
- Start. Press ⇧⌘Space on macOS or Ctrl+Shift+Space on Windows. A small capsule shows the time and your voice level. It never takes focus from the app you are talking about.
- Talk and show. Say what you see. Press ⌃⇧2 / Ctrl+Shift+2 to mark up a region, ⌃⇧1 / Ctrl+Shift+1 to capture the screen, or ⌃⇧Z / Ctrl+Shift+Z for a close-up around the pointer.
- Stop. Press the first shortcut again. One prompt is pasted where you were working and kept on the clipboard. On macOS pasting needs Accessibility; without it the prompt is copied for you to paste.
Any capture shortcut also starts a recording. Every shortcut can be changed in Settings → Shortcuts, and you can hold the key while you talk instead of tapping it.
-
On-device speech. Pick a model once; it downloads and then runs offline.
Model Download Best for Moonshine Base 111 MB Recommended. Fast and accurate English. Moonshine Tiny 30 MB The fastest, for short notes. Parakeet v2 482 MB The most accurate English, with punctuation. Whisper Turbo 574 MB Any of 99 languages, detected automatically. Runs on your graphics card. Parakeet v3 487 MB 25 European languages, detected automatically. Whisper Turbo runs on your graphics card through whisper.cpp: any Vulkan GPU on Windows and Linux, Metal on a Mac, with beam search for fewer wrong words than the processor build. On an RTX 2060 Super it transcribes a 10-second sentence in under a second. Without a usable graphics card it falls back to the processor, and Settings says which one it runs on. The 574 MB model and a small helper program download with it; there is nothing else to install.
The fast engine installs itself. If you use Whisper Turbo on the processor (from an earlier version, say), Pointcast downloads the graphics card version in the background while you are not recording, checks once that your graphics card runs it, and uses it from then on; the Whisper Turbo card in Settings › Speech shows how far it is, with a Cancel. A computer whose graphics card cannot run it keeps the processor version and is not asked again. To turn this off, switch off "Get the faster GPU version by itself" in Settings › Speech. Each prompt in History says how it was transcribed: "Transcribed on your graphics card (RTX 2060 SUPER) in 6 s · 20x real time" means Whisper ran on that card, at 20 times the speed of the recording; "on your processor" means it did not, and a graphics card would be much faster.
-
Mark-up tools. The screen freezes; drag a box, then use pins with labels (P), arrows (A), rectangles (R), the highlighter (H), the pen (D), text (T) and a 2× magnifier (M). Five colors, undo and redo. Return captures, Esc steps back.
-
Pointing marks. Rest the pointer on something for half a second and it gets a ring. The button, link or field under it is outlined and named in the prompt, like
button "Place order". Say "this" or "here" and the spot you pointed at in that moment is marked. Each is a switch in Settings → Capture. -
History and Transcribe again. The last 100 prompts are kept on this computer. Click a picture to mark it up again. Transcribe again writes a recording down afresh, for example after switching to a larger model.
-
Pastes anywhere. Chat apps get the pictures inline; plain text with
[image n]markers is the fallback. A recording with no pictures is plain dictation. You can also copy only, or save aprompt.mdwith PNGs (Settings → Output). -
Pictures as file paths. The prompt names each picture's file, so terminal agents such as Claude Code can open the screenshots.
-
Coding agents over MCP. A local server lets Claude Code, Cursor and Codex fetch the latest prompt themselves. Settings → Output → Agent setup copies the setup for each one.
-
Native on each OS. System Settings layout and system fonts on macOS; Mica, Segoe UI Variable and a Windows 11 Settings layout on Windows. Light and dark follow the system.
-
Private by default. See below.
- Speech is transcribed on this computer. Nothing is uploaded.
- Audio of your newest 10 recordings is kept on this computer so you can transcribe it again, and is deleted with its History entry. Turn this off in Settings → Speech.
- The last 100 prompts, words and pictures, stay on this computer until you delete them in History.
- Pointcast's own windows are left out of every capture.
- Password managers (1Password, Bitwarden, KeePass, Dashlane, LastPass, Keeper, NordPass, Enpass, RoboForm, Proton Pass, Keychain Access, Credential Manager, Windows Security) are painted out before a picture is kept. Add your own apps in Settings → Capture → Never capture.
- The MCP server listens only on this computer, needs this computer's key and refuses web pages.
- Pointcast goes online only to download a speech model and, once a day, to ask GitHub whether a newer version is out. Turn that off in Settings → General.
- Nothing or little is written down. Pointcast says when the microphone is silent or very quiet. Check that it is not muted, that the right one is picked in Settings → Speech → Microphone, and raise its input volume in the system's Sound settings.
- Words go missing when you sit back or the room is noisy. Keep Settings → Speech → Pick up quieter and farther speech on. Turn it off if it ever makes your voice sound wrong.
- A shortcut does nothing. Another app may own it. Settings → Shortcuts says which one is taken; click it and press new keys.
- Transcription is slow. Pointcast says when a model is slower than real time on your computer. Pick Moonshine Base in Settings → Speech.
- Transcribing at the end takes long. The capsule shows a progress bar; Cancel saves what is done and keeps the audio so you can Transcribe again.
- macOS, after an update. macOS asks again for Screen Recording and Accessibility. Turn them back on in System Settings → Privacy & Security, then reopen Pointcast.
- Run the setup again. Menu bar or tray icon → Setup Assistant…
Still stuck? Open an issue.
- macOS: turn off Open at login in Settings, quit Pointcast and drag it from Applications to the
Trash. Its data is in
~/Library/Application Support/dev.pointcast.app. - Windows: Settings → Apps → Pointcast → Uninstall. Tick "Delete the application data" to remove models and history too.
See CONTRIBUTING.md for how the code is laid out, tests and releases, and docs/LOCAL_SETUP.md for the prerequisites on each platform.
Speech runs on sherpa-onnx (Apache-2.0). The models are downloaded at setup, not bundled: Moonshine by Useful Sensors (MIT), Whisper by OpenAI (MIT) and Parakeet TDT by NVIDIA (CC-BY-4.0).
MIT.