How to Use OpenClaw Voice Wake

OpenClaw Voice Wake lets a device listen for trigger words and forward the following command to the active gateway session. The important design detail is that wake words are global. The gateway owns one trigger list, persists it on the gateway host, and broadcasts changes to connected clients and nodes. That keeps a multi-device setup from becoming a mystery where your phone, Mac, and browser each wake on different words.

Know the global model

The docs store wake words at ~/.openclaw/settings/voicewake.json with a triggers array and update timestamp. The protocol exposes voicewake.get and voicewake.set. Empty or messy trigger lists are normalized, and empty lists fall back to defaults. Any node or app UI can edit the list, but the gateway is still the source of truth. That means a wake-word change should appear everywhere once the event broadcast lands.

Route wake words deliberately

Voice Wake also has a routing config. The docs define voicewake.routing.get and voicewake.routing.set, with a default target and optional per-trigger routes. A route target can be the current session, an agent id, or a concrete session key. Use this only when the operator intent is clear. For example, one trigger can wake the current session while another routes to a specific agent. Over-routing is how voice systems become surprising, so keep a small table of trigger words and destinations.

macOS behavior to expect

On macOS, wake-word mode listens for trigger tokens and starts capture after a meaningful pause between the wake word and the next word. The docs describe silence windows, a 120-second hard stop, a short debounce between sessions, and an overlay that shows committed and volatile text. Push-to-talk is a separate path that pauses wake-word listening while held, streams partials into the overlay, and forwards on release. It requires Microphone and Speech permissions, and global shortcut observation may need Accessibility or Input Monitoring approval.

Node behavior to expect

iOS uses the global trigger list for local wake-word detection and can edit wake words through settings. macOS has a local Voice Wake enable/disable toggle. Android currently keeps Voice Wake disabled and uses manual mic capture instead. Those differences are intentional because permissions and background capture vary by platform. Do not assume a trigger that works on the Mac will wake every mobile device; check the platform-specific mode first.

Roll it out safely

Pick trigger words that are easy to say but unlikely to appear in normal conversation. Test one device, then a second device, then routing. Verify that dismissing the overlay does not stop listening permanently; the macOS docs specifically call out hardening around sticky overlays. The OpenClaw Playbook recommends writing the chosen trigger list, route targets, and permission requirements in your workspace notes so Voice Wake feels like an intentional command surface, not a haunted microphone.

Design wake words like permissions

A wake word is not just a convenience phrase; it is a permission boundary for microphone-driven action. Avoid common words, jokes, or phrases that appear in normal household or office speech. If multiple agents share a space, use distinct triggers and route them deliberately. Keep a written list of active triggers because people forget which device is listening and which session will receive a command. When you change the list, test from each enabled platform and confirm the gateway broadcast lands. A clean Voice Wake setup should feel predictable enough that users know exactly what will happen before they speak.

Final verification

Before calling How to Use OpenClaw Voice Wake finished, perform one direct test, one failure test, and one rollback check. The direct test proves the happy path works. The failure test proves the documented guardrail is real, not just assumed. The rollback check tells the next operator how to undo the change without improvising. Save those notes beside the channel, node, or gateway config you changed. OpenClaw gets powerful when agents can act, but it stays trustworthy when every new surface has a small, repeatable verification habit attached to it.