aohp-os/aohp

AOHP, short for Android Open Harness Project, is a research fork of Android being built at the AI Industry Research institute at Tsinghua University. The idea is that AI agents are starting to drive phones on behalf of users, but mobile operating systems still assume a human is the one tapping the screen. AOHP keeps Android's hardware support and app compatibility, then adds system features that make services easier and safer for an agent to call. The README contrasts the two approaches with a shopping example. On stock Android, finding a pair of running shoes under 80 dollars means hopping across Amazon, Temu, eBay, the browser, and Notes, with many taps, copy-paste, and app switches. On AOHP, the operating system offers one task-level entrance, the agent figures out the intent, and the system can talk to several apps in parallel before applying a safety policy and returning a result. The repository lists three main capability areas. Personalized user interaction means the OS can generate task-shaped entrances, find capabilities across apps, and keep memory that survives across app boundaries. Efficient agent interfaces include parallel background execution on virtual displays, structured UI representations for the agent to read, an OS-managed sandbox runtime, unified file handling, and a single subscription channel for events and sensor data. Secure information flow uses typed placeholders for sensitive values like payment cards, a trusted vault for secret-handling steps, and data-flow taint tracking enforced at system boundaries. Tasks flow through five stages: expressing intent, resolving it to system capabilities, picking an execution path through APIs or command-line calls or structured UI operations, passing sensitive actions through a policy layer, and recording memory and audit traces. The team reports tests with OpenClaw agents on ten mobile tasks, where AOHP raised the task completion rate from 43.3 percent to 90 percent while cutting tool calls, duration, token use, and LLM requests by 65 to 74 percent. The project is licensed under Apache 2.0. Source code, build steps, and device images are not yet released, and the repository today holds documentation and a technical report citation.