voice-annotate

Screen capture for coding agents. Draw on your screen while you talk, then paste. Your agent gets the annotated screenshots and a transcript that links each mark to what you were saying.

⌃⌥⌘SSnapshotfreeze the screen and draw on it
⌃⌥⌘RWalkthroughrecord while you keep working
⌃⌥⌘TDictationvoice only

Download How it works

Free and open source. macOS, Windows and Linux.

⌃⌥⌘SSnapshot

A training dashboard with four marks drawn on it: two circles, a box and an arrow

Draw on the screenshot

⌘Vcapture.md what your agent receives

# Screen capture · Snapshot · 0:19

Numbered marks (#1, #2, …) are the user's own annotations, cited where they were drawn.

- frame 1 · 1568×1018 · frame_01.png

## Transcript

[00:00 · frame 1] Okay. Quick review of the run dashboard. This best [#1 circled (301,290)-(450,340) in frame 1] accuracy says 312%. That has to be a units bug. It's probably multiplying by 100 twice. The legend [#2 circled (1241,634)-(1461,784) in frame 1] here sits right on top of the curves. Can we move it below the chart? And this export [#3 box around (290,908)-(425,966) in frame 1] button is floating down here. It should go [#4 arrow from (446,910) pointing at (809,230) in frame 1] up next to the filters.

frame_01.png1568 × 1018, with your marks

Works with and any other agent you can paste into

How it works

⌃⌥⌘S takes a Snapshot: your screen freezes so you can draw on it. ⌃⌥⌘R starts a Walkthrough and keeps recording while you work. The microphone turns on with both.
Circle the number that is wrong, or draw an arrow to where a button should go. Each mark gets a number on the screenshot. ↩ finishes the capture.
A few seconds later your clipboard has the screenshots and a transcript. Paste with ⌘V. Each mark is cited at the word you were saying, with its pixel coordinates.

Shortcuts for

Three capture modes

Snapshot

⌃⌥⌘S

Freeze every display and draw on it while you talk. You can hide the overlay, switch to another window and take the next shot in the same capture.

Walkthrough

⌃⌥⌘R

Record while you keep working. Hold a modifier key to draw marks that fade like a laser pointer. Agents read images, not video, so the recording is turned into up to 30 keyframes.

Dictation

⌃⌥⌘T

Voice only. Speak a prompt and paste it as plain text.

Six tools. Arrow, box, circle, highlighter and pin make numbered marks. The pen draws freehand.
Multiple displays. All of them are captured. In a walkthrough, each keyframe comes from the display your cursor is on.
4K screens. The capture adds a full-resolution crop around your marks, so small text stays readable.
Small pastes. By default the clipboard holds the transcript and a file. Agents that read files open the images when they need them.

Your captures stay on your computer

Transcribed on your device: Speech is transcribed with Whisper. It runs on the Neural Engine on a Mac, and on the CPU or an NVIDIA GPU on Windows and Linux.
Deleted after 14 days: Captures are saved to a folder in your home directory.
No server: voice-annotate needs no account or API key and collects no usage data.

Questions

Which agents does it work with?

Any agent you can paste into. Terminals (Claude Code, Codex) take the transcript text, which lists the image files by path. Chat apps and browsers (Claude, ChatGPT, Cursor) take the attached files. If one target wants something else, the tray menu can copy any recent capture again in another format, including a single composite image.

How fast is the transcription?

On an Apple Silicon Mac a one-minute recording takes about five seconds once the model has warmed up. On Windows and Linux it runs on the CPU with a smaller English model by default; with an NVIDIA GPU and CUDA it uses the large one.

Does it understand languages other than English?

Not yet. Transcription is set to English.

What is different on Windows and Linux?

The modes and the output are the same. The app is a separate implementation with its own shortcuts. On Linux with X11 the capture toolbar is visible in walkthrough keyframes, and the capture tells the agent so. On Wayland only Snapshot and Dictation work, and that mode is new. Details are in the README.

Download voice-annotate

Free and open source under the MIT license.

macOS

Download for macOS

macOS 15 or later, Apple Silicon

Move it to ApplicationsUnzip the download and drag voice-annotate into your Applications folder.
Allow the first launchThe app is not notarized, so macOS blocks it once. Open it, then choose Open Anyway under System Settings → Privacy & Security.
Grant two permissionsAllow Screen Recording when asked and reopen the app, then allow the Microphone.

Windows

Download for Windows

Windows 10 or 11, 64-bit

Run the installerIt installs for your user and needs no administrator rights.
Confirm SmartScreenThe installer is not code-signed, so Windows asks once: More info → Run anyway.
Find it in the trayThe app runs from an icon in the system tray. Its settings can start it at login.

Linux

curl -fsSL https://raw.githubusercontent.com/andropar/voice-annotate/master/python/packaging/install-linux.sh | sh

x86_64, Ubuntu 22.04 or newer · or get the .tar.gz

Run the commandIt installs to ~/.local and adds a menu entry. Nothing needs root.
Start it from your app menuThe app runs from an icon in the system tray.
Check your sessionEverything works on X11. On Wayland only Snapshot and Dictation do, started from the tray or your own shortcuts.

On first start the app downloads a speech model once (about 0.5 GB). After that it works offline.