The Siri you actually wanted.
Sarah is a macOS menu bar voice assistant built with Electron. She listens through a global hotkey, transcribes speech in real time, and either drops the text at your cursor or hands it to an AI agent that can act on whatever app you were just looking at — all without leaving the keyboard.
Pick a single trigger key. Sarah derives the other two modes from modifiers. Default trigger is Right Ctrl, but you can switch it to Right Alt, Right Cmd, F18 / F19, or any safe modifier in Mini Settings → Hotkeys.
| Mode | Hotkey | What happens |
|---|---|---|
| Dictation | Right Ctrl (tap to start, tap to stop) |
Speak → polished text appears at the cursor in any app |
| Command | Right Ctrl + Shift |
Speak → OpenClaw agent acts on the current screen — open apps, write Lark docs, call CLIs |
| Quick Ask | Right Ctrl + Space |
Speak → answer streams into a Spotlight-style overlay you can copy or follow up on |
| Quick Ask fallback | Control + Space |
Same as Quick Ask, no Accessibility/Input Monitoring required |
| Screenshot agent | Cmd + Shift + Space |
Capture the current window into a Command run |
The first screen walks through the checks that matter before a user trusts a desktop voice assistant: Microphone, Accessibility, Input Monitoring, speech provider, agent runtime, OpenClaw Gateway, and local tools. The demo buttons let users try Dictation, Command, or rerun checks without hunting through settings.
A tiny HUD appears while recording. The waveform is driven by live microphone level, the left action cancels the turn, and the right action confirms. In Dictation mode Sarah transcribes, optionally polishes the sentence, and inserts the final text directly into the focused app.
Command mode captures the current app, window title, URL when available, and screenshot path before Sarah's own UI appears. The agent uses that context to act on the thing you were looking at: summarize a page, prepare a Feishu write, call local CLIs, or continue through the OpenClaw Gateway. Tool/lifecycle progress is shown separately from the final answer so the response stays clean.
Quick Ask is for lightweight questions about the current screen. It uses the same answer overlay as Command mode, with Markdown rendering, copy, retry, and follow-up. With OpenClaw Gateway connected, answer deltas stream into the overlay instead of waiting for a one-shot final result.
Click the Sarah icon in the menu bar for the compact home surface. Mini Settings is the deeper control center for permissions, runtime switching, hotkeys, local tools, logs, and first-run recovery.
Capture → Speech → Refine → Act
hotkeys Volcengine Ark Dictation: insert text
mic streaming Doubao Command: OpenClaw agent
screen ctx Apple opt-in filler Quick Ask: streaming overlay
clean-up
Capture — Global hotkeys via uiohook-napi, mic at 16 kHz mono PCM, plus a screenshot + active-app metadata snapped before Sarah's window appears (so the agent sees what you saw, not the Sarah UI).
Speech — Volcengine streaming ASR over WebSocket (gzip-compressed). Volcengine is the supported speech path for dictation. Apple Speech remains an experimental, opt-in local fallback for development (SARAH_ENABLE_APPLE_SPEECH_FALLBACK=true).
Refine — Optional LLM polish via Ark (Doubao Lite). Smart structured polish kicks in for utterances ≥40 chars or ones with filler markers; short clean utterances get the lightweight pass. Result: Typeless-style natural Chinese, not raw transcription.
Act — Three branches share one AgentService queue. Dictation inserts at the cursor via node-insert-text. Command and Quick Ask hit the local OpenClaw Gateway with a minimal prompt + lightweight bootstrap so short turns don't pay the full workspace-context cost.
Local Tools registry — Sarah detects installed CLIs (OpenClaw, Obsidian, Lark/Feishu) and tags each capability as read (auto-allowed), write (approval required), or external (approval required). Approvals persist in encrypted user state and can be revoked from Mini Settings.
Memory — ~/.sarah/sessions/<date>.json captures completed turns. A nightly consolidator at 00:10 generates a daily summary using the same lightweight Gateway path.
curl -sSL https://raw.githubusercontent.com/DHLbigmonster/sarah-desk/main/scripts/install.sh | bashDownloads the latest signed release, extracts, copies to ~/Applications/Sarah.app. About 10 seconds.
First launch: right-click Sarah → Open to bypass Gatekeeper. Then grant Microphone, Input Monitoring, and Accessibility when prompted. The first-run welcome card walks you through it.
GitHub Releases → .zip for arm64 (Apple Silicon) or x64 (Intel) → drag Sarah.app into Applications.
Sarah can run as a menu bar app immediately, but reliable dictation requires Volcengine ASR credentials. Add them in Settings or .env; the local Apple Speech fallback is experimental and disabled by default.
git clone https://github.com/DHLbigmonster/sarah-desk.git
cd sarah-desk
pnpm install
pnpm start| Requirement | For |
|---|---|
| macOS 12+ | Runtime |
| Node.js 22.12+ | Building |
| pnpm | Building (npm install -g pnpm) |
| OpenClaw CLI + Gateway | Command and Quick Ask only — Dictation works without it |
To produce a signed standalone app installed into ~/Applications/Sarah.app:
pnpm run install:appThis is the only supported install path. pnpm run package alone makes an unsigned bundle that loses TCC permissions on every launch.
- Register at volcengine.com
- 全部产品 → 语音技术 → 流式语音识别大模型 → click 立即开通 (free tier exists)
- Create an application, copy APP ID and Access Token
- Open Sarah Settings → Speech card → paste both → Save
Or drop them in .env (copy .env.example first). Sarah reads the credential store first, then falls back to .env.
For natural Chinese output that handles fillers, restarts, and run-ons:
- Sign up at 火山方舟 Ark
- Create an endpoint (Doubao Lite is enough)
- Set
ARK_API_KEYandDICTATION_REFINEMENT_ENDPOINT_IDin Settings or.env
Without refinement, Sarah still works but only normalizes punctuation locally.
brew install openclaw # or: npm install -g openclaw
openclaw onboard
openclaw gateway start
openclaw gateway probeSarah reads ~/.openclaw/openclaw.json, finds the configured port + token, and probes 127.0.0.1:<port>. If the config is missing, the token is missing, or the Gateway is down, Mini Settings tells you the exact next command.
| Variable | Default | Purpose |
|---|---|---|
SARAH_OPENCLAW_GATEWAY_AGENT |
true |
Use Gateway path instead of subprocess CLI |
SARAH_OPENCLAW_WS_AGENT |
true |
Use the native Gateway WebSocket client for streaming deltas and tool progress |
SARAH_OPENCLAW_AGENT_ID |
main |
OpenClaw agent id |
SARAH_OPENCLAW_THINKING |
off |
Thinking level for short turns |
SARAH_OPENCLAW_PROMPT_MODE |
minimal |
minimal / full / none |
SARAH_OPENCLAW_BOOTSTRAP_MODE |
lightweight |
lightweight / full |
SARAH_OPENCLAW_MODEL |
unset | Optional provider/model override |
| Permission | Purpose | Grant via |
|---|---|---|
| Microphone | Voice capture | System Settings → Privacy & Security → Microphone |
| Input Monitoring | Global hotkeys | System Settings → Privacy & Security → Input Monitoring |
| Accessibility | Insert text into other apps | System Settings → Privacy & Security → Accessibility |
Sarah notifies you on first run if any are missing and links straight into the right pane. With stable Apple Development signing, grants persist across reinstalls; only the first install after switching identities needs a re-grant.
src/main/ Electron main process
services/agent/ OpenClaw agent integration + memory + consolidation
services/asr/ Speech-to-text (Volcengine + Apple Speech)
services/config/ Credential store, config resolution
services/keyboard/ Global hotkey handling (uiohook)
services/local-tools/ Local CLI/tool detection, capability metadata, approval store
services/push-to-talk/ Voice mode state machine
services/text-input/ Cursor text insertion
windows/ Window managers (HUD, agent overlay, settings, popover)
src/renderer/
menubar-popover/ Sarah popover (replaces native tray menu)
mini-settings/ Mini Settings control center
src/modules/agent/ Answer overlay panel (markdown + streaming)
src/modules/asr/ Floating voice HUD
src/shared/ IPC channels and shared types
scripts/ Packaging, install, verification scripts
docs/images/ Product screenshots and architecture diagrams (this README)
pnpm start # Launch in dev mode (hot reload)
pnpm typecheck # tsc --noEmit
pnpm lint # ESLint
pnpm test # Vitest run
pnpm run package # Build packaged Sarah.app to out/
pnpm verify:mini # Smoke test the packaged app
pnpm run install:app # Package + copy + sign + install to ~/Applicationspnpm verify:mini checks source wiring, packaged app contents, native modules, and runs a packaged smoke test. Run it after pnpm run package and before pnpm run install:app (install removes the build output).
| Problem | Solution |
|---|---|
openclaw not found |
Install OpenClaw and ensure it's on PATH. which openclaw should resolve. |
| OpenClaw config missing | openclaw onboard |
| Gateway not responding | openclaw gateway start, then openclaw gateway probe |
| OpenClaw answers slow | Keep SARAH_OPENCLAW_GATEWAY_AGENT=true and lightweight defaults. Full prompt/bootstrap modes are dramatically slower. |
| No audio / ASR errors | Check VOLCENGINE_APP_ID and VOLCENGINE_ACCESS_TOKEN in Settings or .env. Volcengine ASR is the supported dictation path. |
| Hotkeys not working | Grant Accessibility + Input Monitoring. Switch trigger key in Mini Settings → Hotkeys. |
| Text not inserting | Grant Accessibility. |
| macOS "cannot verify the developer" | Right-click the app → Open (one-time), or System Settings → Privacy & Security → Open Anyway. |
| Dictation feels weak | Configure Ark refinement (ARK_API_KEY + DICTATION_REFINEMENT_ENDPOINT_ID). Without it Sarah only normalizes punctuation. |
ci.yml— runs on PR/push to main: typecheck, lint, test, verify:minirelease.yml— runs on tag push (v*): builds macOS arm64 ZIP, creates draft Release
To cut a release:
git tag v1.0.0
git push origin v1.0.0Review and publish from Releases.
MIT