Mopu 0.19.1 — Whole-task AI access (preview)

Mopu 0.19.1 makes sustained phone tasks available through standard MCP and a persistent CLI, with one bounded whole-task authorization instead of repeated 200-call handoffs.

Changes

• phone_start, phone_session, and free phone_heartbeat; persistent mopu call --agent … clients need no custom bridge daemon.
• Durable delivery records for all tasks; uncertain outcomes block further input, with no automatic replay.
• Task search, pause, reversible deletion/restore and history; clearer progress, result evidence and AI setup in the console.
• Updated English/Chinese setup and AI integration documentation.

Validation and scope

• 234 Python tests and 18 JavaScript tests passed locally; isolated browser review flow passed.
• 150 synthetic deliveries / 750 calls under one grant; a real G100 completed 206 read-only calls under one grant. No real outgoing bulk-delivery acceptance is claimed.
• Runtime matches the tested installed package. All 61 pre-existing tasks were preserved; job applications remain paused.
• The stable 4 MiB checkpoint transfer implementation is retained. Experimental Android-resident runtime and separate transfer throughput optimization are excluded.

Package version: 0.19.1+stabletransfer. macOS/Linux preview, Python 3.11+. MCP or shell execution is required; not every AI host or phone is independently certified.

Manual takeover, pause, disconnect, restart and grant expiry still stop the grant. Existing tasks do not auto-upgrade or resume. Mopu does not launch models.

See AI client quickstart and acceptance scope. Download the wheel for Python installation or the source archive for setup/integration files; SHA256SUMS covers both assets.

v0.6.0 — Public preview

First public preview of Phone Harness: a local Android workspace shared by humans and their existing AI assistants.

Included

• Graphite desktop/browser interface with live scrcpy video, compact phone controls and visible action feedback.
• Original-agent task queue, checkpoints, reordering and immediate human takeover.
• Physical-screen indicator and explicit on/off requests; compact audio, charging, sample-time and device-activity details.
• Standard MCP bridge plus an optional OpenClaw/Jarvis adapter for original-conversation continuation.
• MIT license, English/Chinese setup guides, contribution/security guidance, issue forms and a no-device UI preview.

Both macOS and Linux GitHub Checks passed on release commit 60b93e5.

Getting started

Use the English README or 中文快速开始.

The source ZIP contains the CLI, macOS launcher source, integrations, tests and UI fixture. The wheel is a Python package, not a ready-made signed macOS app; phone-harness install-app builds the local application using Apple Command Line Tools. Python 3.11+, Node/npm, ADB and owner-authorized USB debugging are required. The viewer remains loopback-only; remote desktop to the trusted host is the intended remote-viewing path.

Validation and limits

Local checks passed: 53 Python tests, 11 action-feedback JavaScript tests, six OpenClaw adapter tests, syntax checks and wheel build. The UI was checked at 1280 × 828 and 850 × 618 with synthetic data. The installed macOS window loaded 0.6.0 and continued the real 720 × 1600 stream with human ownership and physical panel Off.

See VALIDATION.md for the exact historical device-acceptance scope. Real-device evidence is limited to the tested Motorola G100 on Android 15 and the tested Mac. Linux CI is software validation, not Linux phone acceptance. Windows, other phones, multi-day unattended operation and acoustic silence during calls remain unverified. The project provides tools and coordination, not an AI model or a sandbox for unrelated raw ADB programs.

The public screenshot, phone illustration and tasks are synthetic. No private device configuration, task history, device serial, live local token or real-phone screenshot is included. SHA-256 checksums are attached for download integrity.

Feedback

Please use Issues, including host/Android versions, reproduction steps and whether human takeover worked. Review screenshots for private content before attaching them. Security reports can use private vulnerability reporting.