Skip to content

How Studio works ​

Studio is the chat-based builder for creating, testing, and publishing voice agents. You describe the agent, talk to it, and publish, without writing code. This page walks through every panel and the path from idea to a live call. Prefer to work from code? The same agent is available over the API.

1. Start an agent ​

From the welcome screen you have three paths:

  • Describe what you want in plain English and let AI draft the config.
  • Pick a template (sales, support, booking, etc.) as a starting point.
  • Import a saved JSON config if you're cloning a previous agent.

First-time users go through a short onboarding that asks for role, tone, and an opening line.

2. Three panels: chat, config, top bar ​

Inside an agent you see:

  • Chat (left) is your conversation with the AI builder.
  • Config editor (right) is a live, scrollable view of every field on the agent (name, role, language, voice, model, settings).
  • Top bar holds the agent name, Test and Publish buttons, a status pill (dirty / saving / saved), the agent's per-minute rate, and the attached phone number.

Edit in either panel and the other updates.

3. Build by talking to the AI ​

Tell the AI what you want, like "make the greeting friendlier" or "add a step to confirm the email", and it proposes a change card you can approve or reject. Approve a change and it lands in the draft.

The AI also asks inline setup questions when something important is missing.

4. Setup questions block publish ​

Critical fields (role, success criteria, opening line) appear as setup cards. You must answer or skip them before publishing. Each card has a "draft this for me" shortcut if you want AI to fill it in.

5. Test in the browser before going live ​

Click Test to call the agent from your laptop mic. You hear the opening line, the agent transcribes you in real time, and the transcript shows below.

Browser test calls:

  • End automatically after 300 seconds
  • Don't need a phone number attached
  • Are free, with no charge against your credits

6. Publish, then attach a number ​

Publishing pushes the draft to production. If anything is missing, the publish modal shows blockers and lets you fix them inline.

Published agents still need a phone number attached under Numbers before real calls route to them. See the Telephony guide for the BYOC flow.

7. Drafts vs. live: no rollback ​

Changes after publish create a new draft. Hitting Publish again overwrites the live config, and there is no version history. If you need a backup, clone the agent from the agent list before making big changes.

8. Slash commands and JSON edit ​

In chat, try:

CommandWhat it does
/voiceShow and change the voice
/modelPick a different LLM
/test-callOpen the test panel
/publishTrigger the publish flow
/helpList all available shortcuts

The top bar also has a JSON editor for compact configs. Use it to paste a saved config or export the current one for sharing or backup. For the full response shape, see the Agents API reference.

9. Two-factor sign-in ​

Turn on an authenticator app in Settings > Security. Scan the QR code, enter the 6-digit code, and the factor is active.

Once step-up is enabled for your deployment, a session that signed in with a password, magic link or OAuth still has to enter a code before Studio loads. The same applies on the consent screen that connects an MCP client such as Claude. Signing out from the code prompt returns you to the login page.

The API enforces the same rule. With a Studio session token, these routes return 403 with {"error": "mfa_required"} until the session has passed the code check:

  • /v1/billing/* (balance, invoices, top-ups, payment methods, alerts)
  • /v1/providers (your own provider keys)
  • /v1/auth/api-keys, /v1/keys/* (except GET /v1/keys/self), POST /v1/auth/regenerate-key
  • /v1/auth/me/password, /v1/auth/me/export, DELETE /v1/auth/me
  • Carrier, OAuth and connector grants: /v1/telephony/connect|disconnect|validate, /v1/voice-providers/*/connect|disconnect, DELETE /v1/connections/:id, POST /v1/oauth/start, POST /v1/connectors/whatsapp/senders
  • /v1/tool-secrets/:name and the campaign intake writes (PUT|DELETE /v1/campaigns/:id/intake, POST /v1/campaigns/:id/intake/rotate)

Accounts without a verified factor are never blocked. Secret API keys and support (impersonation) sessions are not browser sessions, so the code check does not apply to them. Protect secret keys with scopes and rotation.

What's next ​