screen-grab

A macOS menubar tool that drafts text into whatever field you're focused on — online forms, comment sections, email replies, chat boxes — in your own voice. Two hotkeys, two ways to start: dictate what you want, or let the model compose from what's already on screen.

Setup

You need macOS 13+, Node 20+, and Swift 5.9+ (Command Line Tools is enough).

1. Build

cd screen-grab-brain && npm ci && cd ..
cd screen-grab-mac && swift test && ./scripts/build-app.sh release && cd ..

2. Configure your voice

cp -r voice.example voice

Edit:

• voice/style.md — how you write (tone, sentence length, banned phrases).
• voice/about.md — what you write about (background, proof points, hard rules). Optional but recommended; drafts feel generic without it.
• voice/routing.json — map (app, window title) → few-shot bucket under voice/examples/.

3. Configure the daemon

Create ~/.config/screen-grab/config.json:

{
  "backend": "anthropic-api",
  "model": "claude-opus-4-7",
  "hotkey": "RightCommand",
  "composeHotkey": "RightOption",
  "maxTokens": 600,
  "timeoutMs": 20000,
  "voiceDir": "/absolute/path/to/screen-grab/voice",
  "telemetryPath": "/absolute/path/to/screen-grab/voice/telemetry.jsonl"
}

Backends:

• anthropic-api — needs ANTHROPIC_API_KEY; pay per token.
• claude-code-sdk — uses your Claude Max subscription via claude login; no per-token cost.

For anthropic-api, put the key where launchd can find it:

echo 'ANTHROPIC_API_KEY=sk-ant-...' > .env.local
./screen-grab-mac/scripts/load-env.sh

4. Grant permissions

open screen-grab-mac/build/screen-grab.app

Then toggle screen-grab on in System Settings → Privacy & Security:

• Accessibility — global hotkey + read focused field.
• Input Monitoring — receive keystrokes (macOS Sequoia+).
• Screen Recording — fallback when accessibility can't read the field (browsers, Electron apps).
• Microphone + Speech Recognition — for the Dictate hotkey.

And Keyboard → Dictation must be toggled on system-wide (first enable downloads an on-device model).

Permissions are tied to the app's signature. Every rebuild forces you to re-grant; this is a known limitation of ad-hoc signing. See TROUBLESHOOT.md.

5. Use it

Click in any text field, then:

Hotkey	Field empty	Field already has text
Right Cmd (hold while speaking, release) — Dictate	Polish your dictation into the draft.	Continue the existing text in the direction of your dictation.
Right Option (hold ≥1s) — Compose	Draft from screen context + your voice/about.md.	Continue the existing text using screen context as direction.

A HUD pops up bottom-center with the streaming draft. Enter pastes, Cmd+R regenerates, Esc dismisses.

More

• TROUBLESHOOT.md — common errors and fixes.
• MANUAL_TESTS.md — end-to-end test recipes.
• screen-grab-mac/scripts/doctor.sh — preflight check that prints what's missing.
• tail -f voice/telemetry.jsonl — see every request + accept/edit/regenerate outcome.