OpenThumb — your phone, on your side

Introducing OpenThumb

Today's phone assistants are broken. They wait to be asked. They can summarize a notification, but they can't answer it. They can't touch your apps, they can't remember your rooms, and everything you say takes a round trip to someone's cloud.

OpenThumb is an agent rebuilt for the phone. It works inside the apps you already use, handles the messages you chose to let it handle, and keeps getting better at your work — because it remembers. All of it on the device.

Something arrives
while you are away.

A message, an alert, a thread — from an app you chose. The phone is face-down on the desk and nobody is holding it.

Your rule matches.

An app, a phrase, a cooldown, quiet hours. It wakes on the notifications you picked and ignores the rest — including its own, so it can never trigger itself.

It gets to work,
unattended.

It reads the room the way it learned it — memory written from every reply before — and uses the phone's own shell and files when the job needs real work.

You come back
to it already done.

Nothing left without your Send unless you freed the rule. A ledger shows every reply sent and discarded. The phone used its idle time; you kept yours.

9:14

OpenThumb

KakaoTalk · team room · 09:11

"can you send over yesterday's numbers before the standup?"

rule matched · unattended

OpenThumb · working

Read the thread — this room wants the CSV

Pulled yesterday's export, built the sheet

Replied in the thread it came from

The rule fires whether or not anyone is watching.

Anything you do with your phone,
OpenThumb can do for you.

Assistants rely on integrations and good will. OpenThumb uses your phone the way you do — the actual apps, the actual notifications, the actual files. The only wall is the imagination.

Answering

It replies inside the app a message came from — KakaoTalk, Messages, LINE, Telegram — through the notification's own reply action. No integrations, no relay servers, no copy-paste.

Working unattended

Routines that fire on the notifications you choose, morning briefings before the alarm, overnight sweeps. The phone's idle time becomes your time back.

Files and a real shell

A sandboxed Linux shell on the device. It installs packages, runs scripts and works with real files — because some jobs deserve actual code, not a chat reply.

Full device hands

Screen control through accessibility, plus notifications, clipboard and alarms. You grant the permissions, the agent does the work.

Your phone, not our server

No accounts, no analytics, no server of ours. Bring your own model — Claude, GPT, Gemini under your own key. Everything it knows stays on the device.

Memory that compounds

Every reply teaches it the room it answered in. What this group wants, how that alert should be handled — kept, and brought back next time.

Geekbench-style scorecard

The independent agent.

Siri and Galaxy AI are excellent at being asked things. Measured on what a personal agent can be trusted to do on its own, they are not in the race yet.

OpenThumb0.0

Siri0.0

Galaxy AI0.0

Notification autonomy	OpenThumb	Siri	Galaxy AI
Replies inside the app the message came from	✓	~	~
Acts on notifications you choose, unattended	✓	✗	✗
Per-rule cooldowns, quiet hours, active windows	✓	✗	✗
Learns how each app's conversations work over time	✓	✗	✗
Loop protection — the agent cannot trigger itself	✓	~	~

OpenThumb0.0

Siri0.0

Galaxy AI0.0

On-device privacy	OpenThumb	Siri	Galaxy AI
No account or sign-in required	✓	✗	✗
Conversations stay on the device by default	✓	~	✗
Your model, your key — no vendor cloud in the loop	✓	✗	✗
Memory stored as files you can read and edit	✓	✗	✗
Zero telemetry, ever	✓	~	✗

OpenThumb0.0

Siri0.0

Galaxy AI0.0

Control & trust	OpenThumb	Siri	Galaxy AI
Outbound messages wait for your explicit approval	✓	✗	✗
A ledger of everything sent and discarded	✓	✗	✗
Bring any model — Claude, GPT, Gemini, local	✓	✗	✗
Voice wake — hands-free activation	✗	✓	✓
A real shell and filesystem for complex work	✓	✗	✗

Coverage = supported rows per suite (✓ = 1, ~ = ½, ✗ = 0), compared against Siri (iOS 26) and Galaxy AI (One UI 8) public documentation as of Aug 2026. Not a speed or quality benchmark — it measures what each assistant can be trusted to do on its own. Voice wake is shown losing on purpose: that row is next.

Memory

Memory that knows
your apps.

Every reply teaches it the room it answered in. What this group wants, how that service's alerts should be handled — written to plain files on your device, readable and editable by you, and brought back the next time. The second reply is better than the first.

One file per appKnowledge about one messenger never leaks into how it answers another.

Markdown, not a databaseOpen any memory file and see where a habit came from. Wrong? Edit it.

Local, alwaysMemory never leaves the device. Delete the app and it is gone with it.

# APP · com.kakao.talk — what answering here taught me

## team-room ("스탠업 전 숫자")
- wants the CSV attached, not a summary
- numbers come from /sdcard/exports/yesterday.csv
- tone: short, no emoji, Korean

## alerts from this app
- "결제 완료" → log to money.md, do not reply

Private enough to trust
with your phone.

An agent that answers messages needs your trust more than your awe. OpenThumb earns it structurally, not with promises.

Approval at the edge

By default every outbound reply waits for your Send. Free it per rule when you trust it. A ledger records everything sent and discarded.

No server of ours

There is no OpenThumb account, no cloud, no analytics. Your model key talks to the provider you chose and nothing else.

Your keys, your apps

Claude, GPT, Gemini — your subscription or API key. Nothing is billed through us, because there is no us in the loop.

Three steps to a phone
that works on your side.

Install and bring a modelSideload the APK and add your own key or subscription — Claude, GPT, Gemini. Your account, your choice.
Let it walk you through the gatesThe setup screen checks each permission the agent needs, opens the exact system page, and notices the moment you come back. It ends by writing your first rule from the messengers actually on the phone.
Write a rule"When a message from this app arrives, do that." From then on the phone answers — and remembers how.

Your phone,
on your side.

Something arrives
while you are away.

Your rule matches.

It gets to work,
unattended.

You come back
to it already done.