Voice agents

Build a voice bot anyone can talk to from a link

Step by step: add speech keys, turn on the voice engine on your machine, give a bot a brief, and send a link that lets anyone talk to it — with transcripts and follow-ups.

This is the hands-on companion to how to put an AI voice agent on real phone calls. By the end you'll have a bot that talks to people through a link you send them on WhatsApp, remembers what they said, and updates your records afterwards. Setup takes about ten minutes, all from the Pocket app.

What you need

• A Pocket machine (any size; the voice engine is one small container that uses tens of megabytes of memory when idle).
• At least one speech provider key. One key can do both halves: Cartesia, ElevenLabs and Deepgram each offer speech-to-text and text-to-speech. Add Sarvam if your callers speak Hindi, Hinglish or another Indian language.
• A model. Your project's model is used; small fast models give the snappiest replies.

Step 1: add your speech keys

Open Account → Voice. Each provider has an Add key button and a link to get a key. Paste it and tap Save; Pocket makes one cheap call to the provider to check the key before storing it, sealed, for your account. You'll see Key works. under it.

Step 2: turn on the voice engine

On the same screen, under Voice engine, tap Turn on voice and pick a machine. Pocket installs the engine there and it connects out to Pocket's relay. No ports are opened on your machine, and the audio never passes through a third-party voice platform: it goes caller → relay → your machine → your providers. When the dot turns green and says Online, you're ready.

Tap Talk now to try it yourself: it opens a voice link in your browser, with the project's agent on the other end.

Step 3: give a bot a job

Voice conversations run as whichever bot created the link, with its brief, memory and tools. In Account → Bots, create one and write its brief the way you'd brief a new hire on the phone:

You are Asha, the admissions counsellor for Meridian Classes.
On calls: be warm and brief. Find out the class, the target exam,
and whether they want to come for a demo class.
If they want details, send the fee sheet on WhatsApp.
After every call, add the enquiry to enquiries and refresh
the Admissions dashboard.

Step 4: send a link

Now just ask, in the bot's chat or from a routine: *"Send Riya (+91 98xxx xxxxx) a voice link to talk about the weekend batch she enquired about."* The bot calls send_voice_link with Riya's name and the purpose, gets back a private URL, and sends it with send_whatsapp. The link is valid for seven days and closes when the conversation ends.

When Riya taps it, she sees the bot's name and a Start talking button. The bot already knows who she is and why she's there, because the purpose you gave is part of its instructions for that call.

Step 5: what happens during and after

• During: the bot can use tools mid-conversation: send a WhatsApp or email, schedule a follow-up, or end the call politely. Riya can interrupt at any time, and it stops talking.
• Transcript: every line appears in your project as a run ("Voice call · Riya"), with what she said, what the bot said and any tools it used.
• Follow-up: when the call ends, the bot gets a normal run with the transcript, so it can do the slower work: update the CRM, add a row to the dashboard, or book a callback with create_routine.

Tuning tips

• Latency: most of the delay is the model. A small, fast model (a "flash" or "mini" tier) answers in about a second; big reasoning models can take several.
• Providers: Pocket picks from the keys you have: Deepgram first for listening and Cartesia first for speaking, then the others. For Hindi callers it prefers Sarvam for both; the agent matches the caller's language either way.
• Voice: Pocket picks a default voice for the language from your provider. Choosing a specific voice per bot is coming to the bot screen.
• Keep turns short: the brief should ask for one or two sentences at a time. People interrupt long answers.

That's a working voice bot. When you need it on real phone calls, connect Twilio under Account → Voice and the same bot places calls without changing its brief; the overview post explains when to use which.