AgentVis OS builds the missing layers between LineageOS and AI agents: an operating-system-integrated agent platform where on-device agents can see the screen, reason about it with vision-language models, and act — with the permissions, SELinux policy and product integration that a real OS component requires. Not a sideload; a resident.

A system-level Android agent for screen understanding and input automation.

• Conversation-driven: every task is a conversation, with a full screenshot-annotated transcript of every step the agent took
• See → think → act: screen capture → vision-LLM analysis → input injection, gated by a change detector so the model is only invoked when it matters
• Deep OS integration: privileged system_ext app (screen capture + input injection via platform permissions), live task notifications, and an on-screen interrupt the user can always reach
• Task lifecycle: automatic completion detection ("done" protocol with summaries), stall watchdogs, user interrupts — with the whole state machine persisted per conversation in SQLite

The system-integration counterpart — everything the agent needs outside its own source tree:

• SELinux policy for screen capture and input injection (system_app domain)
• lineage_wd40_* lunch products that integrate the agent into LineageOS builds with zero upstream edits
• Porting guides for three tiers: ROM builders, manual integrators, Magisk users
• Named for the machine's heirloom: where LineageOS carries the lineage, WD-40 keeps the lineage running

Under active development. Repositories are currently private while the platform takes shape.

• AOSP/LineageOS builds on Android 15 — first target: Xiaomi POCO F2 Pro (lmi)
• Any OpenAI-compatible vision-language backend (cloud or self-hosted)
• A porting model where adding a device is data, not surgery: SEPolicies, lunch products and manifests — no framework forks