🎙Voice commands and dictation

Getting started#

1. 1

   Open the command bar

   Press 🎙 Voice on the Canvas toolbar, or Alt+V. The first time, the browser asks for access to the microphone.

2. 2

   Say a command

   The bar shows what it heard, what it understood and what it did. Commands that change the graph can be undone by saying “undo”.

3. 3

   Or type it

   The same commands work typed into the bar, in any browser.

Commands#

English and Portuguese commands both work, whatever the interface language: “Add Sub-Agent” and “Adicionar Sub-Agent” do the same thing.

Naming nodes#

• Say part of a node's name: “auditor” finds “Invoice Auditor” when it's the only match. When two nodes match, the bar lists them.
• The only node of a type can be named by its type: “the Root”.
• “Connect A to B” follows the hierarchy: when only B → A is allowed, that's the link made.
• With one node selected, “Add Skill” puts the new node under it and links them.

Safety#

• “Clear Canvas” always asks for confirmation (say “confirm” or “cancel”), keeps the Root Orchestrator and can be undone.
• “Undo” reverts the last voice command, as long as nothing was edited after it.
• Commands follow your role: a viewer can't edit or run by voice either.

Dictating prompts#

Prompt fields in the inspector (the Root's global behavior rules, each Sub-Agent's behavior rules, descriptions, Reference content) and the console's task box have a 🎙 Dictate button. Text goes in at the caret; press the button again, or Esc, to stop.

Product terms are written in their standard form whatever the recogniser heard: “subagente” → Sub-Agent, “m c p” → MCP, “pre flight” → Pre-Flight, “claude ponto md” → CLAUDE.md.

Languages#

Recognition follows the interface language (English or Portuguese). To dictate English prompts with a Portuguese interface, pick the language in the bar's ⚙ settings.

Privacy: where the audio goes#

• The app never receives audio. The browser's speech recognition turns it into text, and commands are interpreted in the browser by a fixed grammar, with no AI model.
• Browsers with on-device recognition keep the audio on your device: the bar shows On this device. Others (Chrome's default engine, for example) send it to the browser vendor's speech service, shown as Browser speech service.
• Choose On this device only in ⚙ to never use a speech service; voice then stays off where the device can't recognise the language.

Browser support#

Speech recognition needs the Web Speech API. Chrome, Edge and Safari have it; Firefox doesn't, so there the Dictate buttons are hidden and the command bar takes typed commands.

Troubleshooting#