Summary

The phone as a server rather than a client — with everything that follows from what it can then be asked to do.

The capability with no desktop equivalent is the messaging: sending through WhatsApp or Instagram by driving the actual app, and transcribing the voice note that just arrived. Accessibility control is what makes that possible, and also what makes the reach broad — an agent that can tap by visible text can operate most of what is on the screen. The build is labelled pre-release, and how much it can do depends entirely on which Android permissions you grant it.

What it is

An Android app that serves MCP from the phone itself. A desktop agent can read the device's state, look up contacts, place a call, send messages through WhatsApp, Instagram, Messenger or Google Messages, watch notifications, drive apps through accessibility gestures, and transcribe audio.

What you get

device_info returns battery, model, OS, network and memory; get_location returns the best recent fix from the device's providers; phone_alert rings and vibrates to help find the handset
search_contacts finds a number by name, make_call places the call, and send_message sends through WhatsApp, Instagram, Messenger or Google Messages
send_whatsapp_business_message is a contact-safe sender for WhatsApp Business, and whatsapp_automation drives the app through accessibility control for contact selection and typing
notifications monitors and manages what arrives on the phone in real time
App control: list_apps enumerates what is launchable, launch_app opens one by package or name, app_actions runs JSON preset actions, and social_media covers Instagram, YouTube and X interactions
Screen control through accessibility — tap by visible text, content description or coordinates, scroll_screen, search_screen to find a search box and type into it, and global_action for home, back, recents, notifications, quick settings or lock
Audio work: voice_record captures a note, transcribe_audio does speech to text from the microphone or a local file, transcribe_file takes a path directly, and transcribe_whatsapp_audio transcribes the most recent WhatsApp voice note
shell runs commands with safety filters and a timeout, read_file reads from allowed storage paths, http_request makes outbound calls from the phone, flashlight and volume_control cover the hardware, and human_command routes a plain-language instruction across the major tools

Requirements

An Android device running the app from the repository's releases — the current build is marked pre-release. The desktop side connects through the pocket-mcp-bridge package, over LAN or a Tailscale network, so the phone and the client have to be able to reach one another. Messaging, contacts, notifications, location and accessibility control each depend on the corresponding Android permission being granted on the device, which is the real setup step.

Setup effort

One command plus a key — adb install -r PocketMCP-debug.apk, then supply credentials