OpenPocket

Appearance

Project Blueprint ​

This page describes the product direction of OpenPocket as a consumer-ready phone-use agent built around a local, configurable Agent Phone runtime.

Product Vision ​

OpenPocket helps everyday users complete real mobile tasks with an AI agent, while keeping control and sensitive data on their own machine.

The target is not only developer productivity. The core focus is everyday life scenarios such as:

  • shopping
  • entertainment
  • social and messaging workflows
  • repetitive in-app routines

Core Principles ​

Local Agent Phone Runtime ​

OpenPocket executes locally through adb, but the controlled target is configurable:

  • emulator (default)
  • physical-phone (USB/Wi-Fi ADB, ready)
  • android-tv (in progress)
  • cloud (in progress)

Why this matters:

  • users can start quickly with emulator
  • users can switch to a dedicated physical Android phone for production-like behavior
  • no mandatory hosted cloud phone runtime

Human Phone vs Agent Phone ​

OpenPocket uses a strict two-device model:

  • Agent Phone: the controlled execution target (emulator or connected device), treated as clean runtime surface
  • Human Phone: the user’s personal phone, used only for approvals and delegated personal data through Human Auth

This separation reduces accidental leakage from personal daily-use devices into automation runtime.

Local Data + Auditability ​

OpenPocket is not a hosted cloud execution service.

  • device automation runs locally through adb
  • workspace artifacts remain local (sessions, memory, scripts, screenshots)
  • model calls are explicit and configurable; users choose model provider and endpoint
  • human-auth relay state and uploaded artifacts are locally inspectable

Dual Control ​

OpenPocket is designed for both autonomous and manual interaction:

  • Direct control: users can operate the current Agent Phone target themselves
  • Agent control: the agent can operate the same target through planned actions

This enables practical handoff between human and agent in one runtime.

Human-in-the-Loop ​

OpenPocket should always allow users to intervene, inspect, and continue.

Current foundations:

  • observable task lifecycle
  • step-by-step persistence and logs
  • explicit command surfaces (agent, gateway, emulator, script, target)
  • remote auth unblock flow via one-time web approval link and Telegram /auth fallback
  • agentic delegation: runtime stores/describes artifacts, and the agent decides how to apply them

Near-term roadmap:

  • richer target-specific hardening (physical phone first)
  • broader Android TV/cloud deployment completion
  • deeper checkpoint UX for long-running real-world tasks

Experience Layers ​

  1. Runtime Layer: local target runtime (adb) + task loop + persistence.
  2. Control Layer: CLI, Telegram gateway, and web dashboard.
  3. Trust Layer: local storage, auditable sessions, script guardrails, controlled execution.
  4. Collaboration Layer: human and agent share control through execution + approval boundaries.

User Scenarios ​

Shopping ​

  • compare products across multiple apps
  • prepare carts and checkouts with user confirmation before final purchase

Entertainment ​

  • routine content checks, sign-ins, and navigation between media apps
  • repeated daily engagement flows without manual repetition

Social ​

  • draft-assist and interaction setup in social apps
  • structured review before sending or posting

Non-Goals ​

  • not a browser-only desktop automation tool
  • not limited to coding and office productivity tasks
  • not a centralized cloud phone execution service

Summary ​

OpenPocket is evolving into a practical personal phone-use system: local, controllable, auditable, target-configurable, and oriented toward real consumer app workflows.