Everything Pocket does

The whole product, in plain words. Each line is something that works today; the few marked partial ship in part. New features land here first, then in the docs and the app.

Agents

Named bots with a brief, a model, tools and memory, working from your phone.

Bots: A name, a brief, a face, a model, a permission mode (ask, agent, autonomous), a home project, a tool allowlist, notifications.
Memory: Notes a bot keeps across conversations; it adds its own, you edit or delete them.
Runs you can watch: Thoughts, tool calls, files and questions stream to the phone; stop, reply mid-run, quote-reply, follow-ups.
Sub-agents and mentions: Parallel sub-agents and named agents; @name mentions inline; /skill commands.
Skills: Reusable instructions per account and team, with sharing.
Routines and alerts: A prompt on a schedule or once at a time; webhooks that trigger one; alerts on dashboard metrics.
Agents know the app: Every agent carries a brief on where settings live and how calls, campaigns, CRM, knowledge and capacity work.

Code and machines

Real repositories on real machines, reviewed hunk by hunk on a phone.

Projects: From a GitHub repo or from scratch; conversations per project; PRs opened, fixed and merged from the app.
Machines: Your own server over SSH, Pocket cloud machines by the hour, GPU rentals; enrol in one command.
Terminal and desktop: A terminal and a remote desktop on any machine; file browser; preview URLs for dev servers.
MCP servers: Per project and shared per team, from a catalogue of about ninety connectors; secrets sealed.
Artifacts and media: Files the agent saves shown inline and in a gallery; generated images and 3D renders.

Calls

A voice agent on a Twilio number, on the Pocket Link adapter, or on a rooted phone's own SIM.

The engine: End-of-turn detection, barge-in, streaming speech both ways, sixteen speech providers including Sarvam for Indian languages, provider fallback mid-call.
Fast replies: A fast call model, 0.9 seconds from the caller's last word to the agent's first sound; the call prepared while the phone rings.
Knowledge on the line: The agent answers from your documents and says it will check when they have nothing.
CRM on the line: Who the person is before hello; notes and stage changes filed while they are still talking.
Your tools on the line: Owner-defined HTTP tools and remote MCP tools, per bot; WhatsApp, email, follow-ups, transfer to a person.
Voicemail and recordings: Mailboxes and carrier recordings recognised in English, Hindi and Hinglish; hang up or leave your message; retried later.
Quiet lines and limits: Nudges, then goodbye; a longest-call limit; all per project.
Voicemail box: A rooted phone answers calls nobody takes, plays your greeting, and sends you the message as text.
Call lists: Your own enquiries and customers, worked within calling hours with capped retries, follow-up texts, a do-not-call list and every outcome on the lead; spread across every ready phone on the team. Not for promotional calling.
Capacity you can see: About four live calls per core, measured; the engine's size and cost in the app and from the agent.
Voice links: A browser page anyone can talk to, with transcripts and follow-ups.
On-device voicepartial: Pocket Link calls with Android's own recogniser and voice, no keys.
Eyes and hands on the phone: Allow screen capture and a bot sees the phone's screen; with Pocket Link (firmware 0.2) it taps and types too, so it can work your own apps the way you would: a look, one action, another look.

Messages

WhatsApp, email and SMS, read and answered by bots.

WhatsApp: Cloud API numbers and a linked personal WhatsApp; bots read and reply; templates.
Email: Send from a connected address; an inbound mailbox for bots.
SMS: Twilio, or the phone's own SIM through phone control.

CRM and dashboards

Leads, calls and numbers that stay honest because the agents file them.

Leads: Stages, scores with reasons, hot leads, owners; every call, message and note on the record; CSV import.
Call next: The ordered queue of who to call now; one tap dials; the outcome asked on return.
Metrics and leaderboards: Attempts, connects, talk time, outcomes, per owner and team.
Dashboards: Datasets, charts, SQL and guest links, built by agents from a sentence.
Team playbook: Scripts and facts every team bot uses; propose, approve, reject.

Knowledge

What your agents may state as fact, searched by words and by meaning.

Sources: Pasted text, web pages with refresh, PDF, Word and text files; each for all agents or for one bot.
Search: Word matching merged with embeddings on your own key; nothing found means nothing found, not a guess.

Phone control (Android)

The handset as a device agents can drive.

Pocket as the phone app: Calls placed, answered, held, merged and swapped; call log; contacts; urgent-caller rules; SMS; apps driven by agents.
Status that cannot be swiped away: Calls, agents, reviews and results in one notification with actions; a Live Update chip on Android 16 where allowed.
Hold to talk: A composer shaped like WhatsApp: a plus, the text field, a mic that becomes send.
Root mode: Hear and speak on calls with no adapter; the voicemail box.

Teams

Roles, shared resources, and the team's phones as a pool.

Orgs and roles: Owner, admin, billing, member, viewer; invites; SSO and domain claims; audit log; spend caps; usage per member.
Shared resources: MCP servers, skills, agents, knowledge and model keys at the team level.
Workers: Every member's phone works the campaign list; pause and reassign from chat; health alerts to the lead.

Models and keys

Your keys, Pocket's credits when you have none, and the right model for each job.

Bring your own keys: Twenty-plus providers and custom endpoints; a catalogue of what each key unlocks.
A fast model for calls: Separate from the chat model, routed to the fastest host.
Provider health: Credit and outage alerts; speech-provider fallback.

Billing

Free with your own server and keys; plans and credits for the rest.

Plans: Free, Pro, Team; limits on bots, channels, runs, call minutes and machines.
Credits: Top-ups by card, auto top-up, postpaid invoices for teams; replayed webhooks never double-credit.
Machines at cost: Google's live regional rate; GPU rentals and paid tools re-billed with a stated margin.

Security and privacy

Every door checked before every release.

Sign-in and tenancy: Google sign-in and SSO; personal and team tenants; one access check for every resource; user B never sees user A's data.
Secrets sealed: Model keys, MCP headers, tool headers and device tokens encrypted with a master key and never returned.
Verified callers: Runner, cron and engine doors take their own secrets; payment, GitHub and WhatsApp webhooks verify signatures; replays ignored.
Tool calls on a leash: Everything a call's agent runs goes through the server and the bot's allowlist; nothing said on the line reaches a tool it was not given.
The release gate: Every API route swept anonymously with every method, forged webhooks, the tenant line, the call engine's rules: all green before an image is built.

Want the detail? The docs cover each area, the blog has the measurements, and pricing says what costs what.