DeepseekHarness-Android

What it solves

Running DeepSeek Harness (dsh) on a phone usually means Termux, proot, or root, plus fighting Android’s background-process limits. DeepseekHarness-Android compiles Node 24 into a native Android shared library and boots dsh via JNI inside the app’s foreground service, so a single APK is enough:

• No root, no Termux: the engine lives in the app process, avoiding the Android 12+ phantom-process restrictions.
• Device capability bridge: all phone actions are implemented in Kotlin and exposed to dsh over a private socket channel — file search/organize/archive with PDF/Word/Excel/PPT text extraction, accessibility-based screen reading, tap, swipe and text input, system settings (brightness, volume, DND, alarms, Wi-Fi, Bluetooth, location), notification read/reply and media control.
• Automation triggers: scheduled runs, or tasks fired on Wi-Fi/Bluetooth connection, charging, arriving at a location, or receiving a notification.
• No fork of dsh: the app talks to dsh through its SDK protocol, with Android adaptation kept in a profile, plugins and small patches, so skills, MCP, sub-agents and web search keep working. Besides DeepSeek, many OpenAI-compatible providers can be added.
• Optional Shizuku: unlocks extra system toggles and boot-start over wireless debugging.

Who it’s for

• Users who want to try or drive a device-operating agent on a phone without setting up Termux or root.
• Developers who need DeepSeek Harness as a mobile automation base for scheduled jobs, conditional triggers, and notification/file handling.
• Engineers curious about the “embedded Node native library + capability bridge + trigger hub” architecture for Android agents; the repo ships architecture notes, an engine contract, and full build steps.

Requirements: Android 10+, arm64-v8a, ~110 MB APK, cold start around 5 seconds; other apps cannot be operated while the device is locked; third-party models are limited to OpenAI-compatible endpoints.