Calls: three ways onto the line

An ordinary Android app cannot hear or speak into a phone call; Android keeps that for the phone app and the system. Pocket gets a bot onto a call three ways, and everything after the call (outcome, talk time, transcript, the next step) works the same whichever you use.

• Pocket Link (your own number, no root, no Twilio): a small adapter plugs into the phone and carries the call's audio to the bot. The phone's own SIM makes and takes the calls, with the app closed. Set it up.
• A rooted phone (your own number, no hardware): with root, Pocket hears and speaks on calls directly, and can run a voicemail box. Phone control → Root mode.
• Twilio (a number with no hardware): connect your Twilio account under Account → Voice → Phone calls and any bot can ring a number from it, talk, and file the outcome. Twilio bills the minutes to your account.

Connecting Twilio

Under Account → Voice → Phone calls enter the Account SID (AC…), an API key SID (SK…) with its secret, and the number. Find them in the Twilio console. A *verified caller ID* works for calling out, but nobody can call or text it; buy a Twilio number if the bot should also answer. Inbound calls to the number go to the bot you mark Answer my calls in its set-up.

Speech providers and the voice engine

Account → Voice → Speech providers holds the keys the bot listens and speaks with. Only streaming providers are offered, because a call's latency is the sum of its parts: Deepgram and Cartesia for both listening and voices; AssemblyAI and Azure AI Speech for listening; ElevenLabs, Inworld and Fish Audio for voices. Hindi and Indian English are covered by Deepgram, Cartesia, Azure and ElevenLabs. The bot listens with speech-to-text, thinks with a fast call model, and answers in the voice you pick; if a provider fails mid-call the engine falls over to the next one you have a key for. The bot answers in the caller's language: English, Hindi, Hinglish, whatever is spoken. Talk now on the same screen lets you try the voice in the app before a real call. The engine is Pocket's, shared, with room for about four calls at once; a team that needs more runs it on a machine of its own.

What every call leaves behind

When the call ends, the bot's run continues as a normal conversation with every tool: it files the outcome (connected, no answer, busy, voicemail, failed), the talk time and its notes on the CRM lead, sends the WhatsApp it promised, schedules the follow-up, and reports to you. Calls you make yourself from Call next ask for the outcome when you come back to the app, so the numbers stay honest either way.