apptah builds agents and apps that plan, act, and hand back finished results —
files written, posts published, answers grounded in fact — while a human holds the
approve button. And where plain arithmetic beats a model, it ships the arithmetic.
Not a reel of prototypes — a working ledger. Each entry below is in someone's hands today:
on the App Store, live on the web, deployed on private hardware, or free to fork.
012026 — present/macOS & WindowsFlagship
OpenCowork
An agent that finishes the work.
State an outcome, approve the plan, and get finished files back — no terminal,
no code. A desktop agent that does the task instead of talking about it.
Plans, then acts — reads and writes real files, produces finished documents, and reports back exactly what changed
14+ interchangeable AI providers behind one interface; new abilities load without a rebuild
Safe by design — it touches only the folders you grant, and every change is shown for approval before it lands
300+ automated tests on the core decision logic; ships as one self-contained installer
fig. 01 — an outcome requested, a plan in motion, files on the way back.
022025 — present/WebLive in production
Autopost
A social channel that runs itself.
It finds what's moving, rewrites it in your voice, and publishes on schedule —
while a human keeps the approve, edit and reject buttons.
Learns your voice from what you've already written, then drafts, schedules and posts automatically
Every client isolated, with daily spending caps and usage logs so costs stay predictable
Publishing that can't double-fire — a post never goes out twice, never fails silently, and stuck jobs recover on their own
fig. 04 — the repository: a data-centre model, fitted into a pocket.
052026 — present/iOS
Champions Advisor
AI that cannot make things up.
Screenshot the team preview and it tells you what to bring and which move
to use — with the maths already done.
The advice is arithmetic, not a guess — damage, speed and matchup calculations run as a deterministic engine under 89 automated tests
Reads the screen on-device; the AI may only explain results the calculator has already proved
Run as a business — paid subscriptions with per-user spend caps
Logic tests89
PerceptionOn-device
Model's roleExplain only
fig. 05 — six against six, the counter-play already costed.
062026 — present/iOSOn the App Store
HKeTravel
Knowing when not to use AI.
The genuinely fastest way across Hong Kong, and when your bus actually
arrives — deliberately no AI, because timetable maths gives the more reliable answer.
Every mode in one route — bus, minibus, rail and light rail, ranked by real door-to-door time with live arrivals
The fare shown is the fare paid — real interchange discounts included
Fast and honest — start-up cut from 7 seconds to 1, works offline, published in four languages
fig. 09 — the memory an agent keeps between conversations.
102025 — present/Client deliveryDelivered
COMFi
A whole storefront, delivered.
The complete online store for a Hong Kong brand selling into nine-plus Asian
markets — payments, order emails, alerts and content management, handed over working.
Built as custom modules — payments, order emails, web push and chat alerts, plus a lightweight content manager
Caught a pricing fault before it charged anyone — a bug that would have billed customers 100× the correct price, found and fixed first
Three languages, automated tests, hardened payments — live on the client's own domain
fig. 10 — the client's shop, open in three languages.
Also in the ledger — an on-device translation app that never calls an API ·
a cartoon image generator built for educators · and the smaller ships between.
Twelve built; ten shown.
№ 02 — The method
How the studio works.
Five rules, applied to every entry above. They are why the ledger reads the way it does.
1.Finish, don't chat.A product here must hand something back — files written, posts published,
answers grounded. Conversation is overhead, not output.
2.A human holds the approve button.Drafts, plans and file changes stop at a person: approve, edit, reject.
Autonomy without a checkpoint is a liability, not a feature.
3.Private by default.On-device and self-hosted first. Confidential material never leaves the
building — that's the starting point, not a pricing tier.
4.AI only where it wins.Where arithmetic beats a model, the arithmetic ships — and the model is
allowed to explain, never to decide.
5.Shipped is the standard.App Store listings, live services, paying customers, open licences.
The ledger counts what shipped, not what was demoed.
№ 03 — In the open
Two of them are free to take apart.
Shipping in the open is the strongest proof there is. No activation, no cloud dependency, no catch.
PrivateAgent
The 35-billion-parameter model on an iPhone, with the engine that streams it from
storage and compresses its memory. Everything on-device; nothing phones home.
The memory layer for personal AI agents — keyword, meaning and structured-fact recall,
combined, offline. Drop it beside an agent and it stops re-reading everything.
The studio takes product engineering, private AI deployments of its own tools,
and the occasional impossible brief. One address reaches a person — not a funnel.
OpenCowork
