Voice

Talk to your agents by voice — hands-free agent control over the phone or from your browser.

Overview#

Voice gives you a real-time conversational interface to chat.dev. Instead of typing commands, you speak naturally and the system handles the rest: understanding what you said, executing the right action, and speaking the results back to you.

Two ways to use it:

• Phone — call +1 (856) 242-8338 (+1-856-CHATDEV).
• Browser — click the mic button next to the send button in the chat view.

Both surfaces connect to the same AI assistant with simple, name-based commands — just like texting.

Phone#

1. Sign up first. You need a chat.dev account linked to your phone number. If you haven't already, text the chat.dev SMS number to create an account.
2. Call +1 (856) 242-8338. The system recognizes your phone number and connects you to the voice assistant.
3. Speak naturally. Say things like:
   • "List my agents."
   • "Create an agent called Builder."
   • "Ask Builder to build me a REST API on port 4000."
   • "What's my balance?"
4. The assistant acts immediately. No confirmation prompts — it just does it.

If you call from an unregistered number, you'll hear a message asking you to sign up via text first.

Browser#

1. Open the chat view for any agent on chat.dev.
2. Click the mic button (next to the send button). Your browser will ask for microphone permission.
3. Speak your command. The assistant listens, understands, and responds with voice.
4. Click the mic button again to end the voice session.

While voice is active, you'll see status indicators: Listening... when the system is waiting for you to speak, and the assistant will respond through your speakers.

How it works#

Voice routes each request from the spoken context and the agent you name:

• Name an agent when you want to send work to a specific one: "Ask Builder to run the tests."
• Browser voice starts scoped to the agent page you opened. If you start voice from Builder's chat view, Builder is the obvious target for coding tasks.
• If you only have one agent, it can be used automatically.
• If you don't have any agents, just describe what you want built — the assistant creates one for you.
• Direct live conversations are opt-in. The Direct Speak toolpack enables commands such as talking directly to a named agent. Without that toolpack, voice can still create, start, stop, inspect, and send tasks to agents, but it does not enter a persistent direct-agent conversation.

When you send a task by voice, it appears in the chat view and terminal, just like a typed message. When the agent responds, you hear it spoken back to you.

If a coding agent gets blocked on an interactive choice, voice speaks the input-needed prompt. You can answer by voice, by SMS, or in the terminal depending on the situation.

What you can do#

• Create agents — "Create an agent called Builder."
• Send tasks — "Ask Builder to build a todo app with authentication."
• Direct speak — "Talk to Builder." when the Direct Speak toolpack is enabled.
• Start/stop/restart — "Stop Builder.", "Restart my agent."
• Check status — "What's Builder doing?"
• Get output — "What did my agent say?"
• Check balance — "How much credit do I have?"
• Rename/delete — "Rename Builder to Frontend.", "Delete my test agent."

Tasks sent to agents appear in the chat and terminal views. Meta-commands (like "list my agents" or "check my balance") are answered by voice only.

Tips#

• Speak clearly and concisely. The system uses voice activity detection — pause briefly after finishing a thought.
• You can interrupt. If the assistant is speaking and you start talking, it stops and listens.
• Use agent names. Say "send Builder a task" or just describe what you want if you only have one agent.
• Responses flow back. When your agent finishes a task, the assistant tells you what happened — no need to check the dashboard.
• Answer input prompts directly. If the agent is waiting for a numbered choice or confirmation, voice will announce that input is needed and read the options when they are available.
• Background thinking. If you pause after a response, the assistant may follow up with a deeper thought or suggestion.

Voice profiles for agents#

Each agent has a voice profile used when it speaks on calls and in browser voice:

• voiceName — the realtime voice: marin or cedar (lowercase). When you create agents without specifying one, chat.dev alternates voices so agents in the same account sound distinct.
• voiceAccent — a speaking accent flavor: English, German, French, or Italian (any capitalization is accepted on the API).

Both can be set at creation (POST /api/agents) or updated later (PATCH /api/agents/:id).