Changelog

1.1.0 12 September 2026

First published release. What it ships with:

• The agent loop. A task in plain words, read the screen, one action, read again. It can open apps, tap, long-press, swipe, scroll, type, and press back, home or recents, and it keeps a checklist across the steps.
• Rules, approval mode and plan mode. Three layers checked before every action, covered on the overview.
• Chats. Append-only, so a follow-up resumes the real exchange and a run that was killed can be picked back up. Long-press a message to fork the chat there.
• A second, cheaper model for the mechanical steps, with the chosen model kept for planning, recovery and anything you say.
• Lua scripts the agent can write and run on the phone, for the steps that have to be decided as they happen. Every action inside one is checked and logged like any other.
• Saved scripts, callable by name and runnable by hand from a form, with no model in the loop.
• Backup and restore. One file carrying chats, scripts, saved tasks, playbooks, rules, memory, servers and settings, so a reinstall or a new phone is not a fresh start. Importing merges, so it never deletes what was only on the phone you opened it on.
• Saved tasks, for the ones you run often.
• Playbooks, per-app notes that load only when that app is in front.
• MCP servers over streamable HTTP, with a bearer token or none.
• A status indicator, for watching a run and stopping it from inside another app, with questions reaching your notifications too.
• Diagnostics export, which leaves out your API key, server tokens and anything typed into a password field.

The app is called Android Agent from this release. Earlier builds carried the repository slug on the launcher and in the accessibility list, and the accessibility description named an API the app has never required.

Fixed: an accessibility service that Android rebound could be left reported as off while it was running, which stopped runs from starting at all.

Download 1.1.0