To dictate to OpenClaw on a Mac, click into the chat composer you already use, whether that is the dashboard chat, Quick Chat, the terminal TUI or a messaging app, and speak with your Speakmac shortcut. Read the text before you press Return. OpenClaw acts on your files, accounts and messages, so a wrongly heard name or path costs more than it would in an ordinary chat.

OpenClaw is the open-source personal AI assistant started by Peter Steinberger in November 2025. You may know it by an earlier name, Clawdbot or Moltbot. It was renamed OpenClaw in January 2026 and is now looked after by the OpenClaw Foundation. A Gateway runs on your machine, connects to the model you choose, and answers you in chat apps and its own dashboard. There is no Speakmac integration: Speakmac supplies text, and OpenClaw and its model handle everything after you send.

Where do you type instructions to OpenClaw on a Mac?

The dictation workflow is the same in each of these:

  • Dashboard chat. The macOS menu bar app embeds the Gateway's Control UI. In a browser it is served at http://127.0.0.1:18789/ by default. Return sends and Shift-Return adds a new line.
  • Quick Chat. Press Option-Space for a floating composer that sends to the agent's main session.
  • Terminal. Run openclaw tui. Enter sends, and Shift+Enter or Ctrl+J inserts a new line (TUI docs).
  • Messaging apps. WhatsApp, Telegram, iMessage or Slack on your Mac, if you have connected those channels.

For long briefs, the dashboard composer is the easiest place to review before sending.

How do you dictate a task to OpenClaw?

  1. Install Speakmac and finish the microphone and Accessibility setup. Add the names OpenClaw will see often, such as contacts, project names and skill names, under Words & Text → Custom words.
  2. Open the OpenClaw surface you want and choose the right agent and session.
  3. Click the composer so it has keyboard focus, then use your Speakmac Push to Talk or Hands-Free shortcut.
  4. Say the outcome, what the agent may and may not touch, and when it should stop and ask you. Then stop dictating.
  5. Fix names and paths. Type exact values such as $release_notes; typing $ in the Control UI composer searches your skills. Send only when the brief is right.

If the agent heads the wrong way, typing /stop aborts the run.

What is a useful voice brief for OpenClaw?

Go through my unread email from this week and draft replies to anything about the October invoices. Do not send anything. Put the drafts in a list in this chat with one line on why each needs a reply, and flag anything that mentions a refund for me to handle myself.

Every weekday at eight, check the build status for the website project and message me on Telegram only if something failed. Include the failing step and the last commit message. If you are not sure whether something counts as a failure, ask me instead of retrying.

These are example briefs. We did not run them against a live OpenClaw agent.

Instruction files are where long dictation helps most. OpenClaw loads AGENTS.md and SOUL.md from its workspace at the start of every session, and each skill is a SKILL.md file in a folder. Dictate a rough draft in your editor, then cut it down: large bootstrap files are truncated when injected (the default limit is 20,000 characters per file), and every rule you add costs context.

What should you type instead?

Speak itType or paste it
The goal and why it mattersPaths, URLs, email addresses
Limits: what not to send, delete or buySkill names, slash commands, cron times
When to stop and ask youError output and config values
Rules for AGENTS.md and skillsAnything secret

Never dictate secrets. OpenClaw's own security docs say session transcripts can contain pasted secrets and are readable by anyone with access to the disk. When OpenClaw needs a credential, the TUI provides a masked prompt that keeps it out of chat history. Keep the scope narrow too. The workspace is not a sandbox unless you enable one, third-party skills should be treated as untrusted code, and the dashboard is an admin surface that should not be exposed to the internet. Speakmac transcribes on your Mac, but once you press send, your model provider's policy covers the text.

What if text does not appear?

Click back into the composer and try a short sentence. If Speakmac produced a transcript but it did not appear in the composer, use Copy Last Transcript in Speakmac and paste it. If no transcript appeared at all, check your microphone and model first; that is a separate problem from focus. The Help Center covers Accessibility permission and shortcut clashes. OpenClaw's push-to-talk uses the right Option key, so do not assign that key to Speakmac as well.

Does OpenClaw have its own voice input?

Yes, and for short requests it may be all you need. On macOS 26.2 or later, the app offers Voice Wake and push-to-talk. Hold the right Option key, speak, and release; the text is sent after a short delay. Quick Chat has a microphone button, Talk mode holds a spoken conversation, and voice notes sent from WhatsApp or Telegram are transcribed by your configured provider or a local Whisper-style tool.

Wake-word mode sends after a pause in your speech, and the docs say push-to-talk, Talk mode and Quick Chat dictation may use Apple's network speech services. These features also work only inside OpenClaw. A system-wide dictation app is useful when you write long briefs or edit AGENTS.md and skill files, want custom words for names, want to review before anything is sent, or want the same shortcut in your editor, terminal and chat apps.

Try Speakmac for Mac, or see the same workflow for Claude Code and Codex.

Related posts

Back to top