Machines
Calls on your machine
Run Pocket's call engine on a machine you rent or own, so your calls are carried by its cores. One switch; nothing exposed to the internet.
Calls with natural voices (a Twilio number, a voice link, a Pocket Link or rooted phone with cloud voices) run on Pocket's call engine: it hears the caller, decides when they are done, thinks with a fast model, speaks, and stops the moment it is talked over. Everyone shares one small engine, which carries about four live calls at a time. A team that needs more lines runs the same engine on a machine of its own.
Turning it on
- Open the machine under Machines (a Pocket-hosted one, your own server, or one in your cloud account) and turn on Calls on this machine. Or ask a bot: "run our calls on the standard machine".
- Pocket installs its engine on the machine (a few minutes the first time) and keeps it running through the machine's agent; the machine then holds one connection to Pocket. Nothing listens on the machine, so it needs no public address and no open port.
- From the next call on, the machine takes your calls. The switch shows Connected · 2 of 8 lines in use when it is up.
- Set the machine to always on. A machine that has gone to sleep cannot take a call, and waking takes longer than a ring.
What runs where
- On your machine: the audio in and out, hearing when the caller is done, barge-in, the voice; speech providers (Deepgram, Cartesia, ElevenLabs…) spoken to on your keys; the model spoken to directly when you bring your own key.
- On Pocket: setting the call up, the bot's brief, the caller's CRM facts, every tool used on the call (WhatsApp, email, calendar, transfer), the model on Pocket credits, the end-of-call report and transcript, the after-call run, billing.
Speech keys are handed to the machine for each call and held in memory only while the call runs. The machine's engine authenticates with a token of its own that can fetch your account's calls and nothing else; turning the switch off ends it.
How many lines
About four live calls per vCPU before audio detection starts to lag. Starter (2 vCPU, shared cores) carries about 4; Standard (2 vCPU) about 8; Pro (4 vCPU) about 16; your own 8-core server about 32. Past that, calls above the limit go to Pocket's shared engine rather than failing. The bot's call_capacity tool and the machine's switch both show the number.
When the machine is away
If the machine is asleep, offline or full at the moment a call comes in, the call runs on Pocket's shared engine instead; nothing is missed, and the machine takes the next one when it is back. Each call's run says which engine carried it.
What it costs
A Pocket-hosted machine is metered by the hour at Google's price plus 30% while it runs, always on for calls: Starter about ₹1,100 a month, Standard about ₹4,400, Pro about ₹8,800; the day's total shows under Account → Plan. Your own server or cloud account costs nothing from Pocket. The model is credits or your key, speech and the phone line are on your provider accounts, as on the shared engine.