Layer 01 · Toolkit in the forge version — in the forge

agenticdroid

An agentic AI toolkit for phone control — your agent sees your Android phone's screen and drives it.

Binds to a field of work — phone control. Android, where full control actually exists.
Get the Forge Pass Back to the explorer
Ground Truth · How it aims

Every tap lands on the map, not on a guess

  • ▸UI-map ground truth — targets come from the interface's own accessibility map, so a tap is a coordinate with a name, never a pixel prayed over.
  • ▸Watch, then take over — you do a task once; the playbook is written from your demo and run without re-asking.
  • ▸One stop ends everything — a single stop halts all input at once; your hand on the phone outranks the agent, always.

Schematic: the phone above is a drawing of the method — a screen read as a map of named targets, with the path between them traced — not a screenshot of a session.

What it is

agenticdroid, in full

Born from real phone-control work: your agent connects to an Android phone over wireless debugging and runs it — seeing the screen, tapping with UI-map precision, organizing apps, clearing clutter, working through tasks — over scrcpy and VNC channels reachable anywhere via Tailscale. It learns by watching you do a task once, then takes it over, with hard stop rules and no-touch zones built in. Android is where full control exists: stock iPhones allow view-only access at best, and agenticdroid doesn't pretend otherwise.

Two channels, one discipline: scrcpy over wireless debugging for the working channel and VNC as the fallback, reachable anywhere via Tailscale. Sessions run under written rules — read the last session note before touching anything, write a note on every pause, and never pursue a tap after a failed gesture without re-checking the map first.

The no-touch zones are part of the product, not fine print: no purchases, no permission dialogs, no connectivity switches, no unnamed dialogs — and folders or apps are never deleted without explicit approval of the verified target.

Capabilities

Wireless debugging + scrcpy controlVNC fallback channelReachable anywhere via TailscaleUI-map ground truth for every tapWatch-then-take-over playbooksApp, folder & cleanup runsHard stop rules & no-touch zonesLearns from your corrections

Status — honest label

In the forge — techniques proven on a real phone

Proven in live sessions

  • ▸Real work, real device — the control techniques were developed driving an actual Android phone through actual jobs: app and folder organization, cleanup runs, owner watching
  • ▸Watch-then-take-over playbooks exist and run — demonstrated once by the owner, executed by the agent after
  • ▸The stop rules were earned, not imagined — every hard rule in the toolkit answers to a moment in a real session

The honest boundary

  • ▸Android only, by physics — full control exists where the platform allows it; stock iPhones offer view-only access at best, and agenticdroid says so instead of pretending otherwise
  • ▸The packaged toolkit is still being forged — sessions and playbooks today; the self-contained repo is the in-the-forge work

Keep going

agenticdroid in the family

agentpreneur

The business-building sibling: agentpreneur spins up the venture; agenticdroid runs the phone the venture lives on.

Meet the sibling →

Lingo — built in

The same six prefixes steer the phone work: sch: for when, mem: for what it remembers about your apps and folders.

See the plugin →

The full explorer

All four toolkits and both plugins, grouped the way the forge files them: by what they bind to.

Browse the catalog →

Product agenticdroid Layer Toolkit Status in the forge Access Repo access with the Forge Pass