Pinned Tweet
I gave an AI agent my iPhone.
No Mac. No simulator. Just a USB cable and a laptop.
It tapped through my app, checked every screen, and even posted a tweet for me. 🧵
قاعد افكر جدياً اني انسخ T3 Code في نسخة خاصة فيني أحدثها شهرياً بالتحديثات الي ينزلونها
احتاج استخدم TapWright مع T3 من تلفوني عشان أتحكم بآيفوني الي فالبيت من آيفوني الي استخدمه يومياً
2
254
Tapwright 1.9.0 is out: the Token Update.
We just cut the token cost of controlling a real iPhone with AI agents by 31% — without removing a single feature.
Here’s the counterintuitive lesson we learned building it: cutting tokens isn't about shorter tool outputs. It’s about cutting turns.
Tapwright lets agents (Claude Code, Codex, Cursor, or any MCP client) see and drive a real iPhone from macOS, Windows, or Linux.
To make 1.9.0 faster and cheaper, we benchmarked 21 real-world tasks across Haiku, Sonnet, and Opus on a physical iPhone before writing a line of code—filling forms, digging through long lists, changing settings, and filing bug reports.
The data revealed a massive bottleneck: Tapwright’s responses were only a few thousand tokens per task. The real bill came from turn count, because every new turn forces the model to re-read the entire conversation history.
So we engineered 1.9.0 to eliminate wasted turns:
1. Automatic route memoryWhen an agent navigates to a screen, Tapwright quietly remembers the path (IDs only—never typed inputs). On the next session, the first response tells the agent what's one call away, letting drive {"route": ...} jump straight there in a single step. Even search-then-tap flows are safely cached.
2. Guidance where agents actually lookOur benchmarks showed agents rarely open external skill files. So we moved the core rules directly into the tool descriptions and responses: batch known steps, target by ID, and only screenshot to check visuals. One short hint per answer, shown once.
3. The iOS keyboard stops breaking flowsResponses now flag what the on-screen keyboard is covering. If an agent taps a hidden "Save" button, Tapwright automatically dismisses the keyboard first—eliminating a failure loop that used to burn a turn on almost every multi-line form.
4. Zero-latency tool loadingPreviously, every session burned a full turn just discovering Tapwright's tools. Core tools now load immediately with the session, dropping setup turns across our benchmark from 29 down to 2.
5. Fewer dead endsSmarter assertions that accept what models naturally write, graceful fallbacks when OCR isn't installed, app-only logs, and self-summarizing bug reports.
The benchmark results (same 21 tasks, same phone, same 3 models):
31% fewer total tokens
30% fewer model turns
26% lower API cost
-48% tokens on form-filling tasks
-52% tokens on repeat screen visits
Opus’s explore-and-file-a-bug run dropped from 40 calls to 19
Nothing was taken away. Every 1.8.5 call still works out of the box, and --no-hints restores the exact 1.8.5 behavior if you want it.
Upgrade today and let your agents build faster for less.
I gave an AI agent my iPhone.
No Mac. No simulator. Just a USB cable and a laptop.
It tapped through my app, checked every screen, and even posted a tweet for me. 🧵
2
303
Try it: pip install -U tapwright github.com/BO5AMIS/tapwright
48
Can we please get a way to @ a specific thread or agent or even sub agent in chat in T3 Code? It’s becoming tedious copying thread ids or dragging and dropping. Specially on mobile.
I usually have multiple chat threads that are part of a larger team so this will help a lot.
@jullerino @theo
159
مب معقول شلون Astra تعيس حتى بعد الإصدار الجديد في متابعة الأوامر
وابسط الأشياء يعقدها ويتفلسف ويهذر وايد
حتى شغله في Blender و unreal engine أسوء من Opus
شكله صج لازم نلغي الاشتراك
1
336
It doesn’t guess from pixels. It reads every button and label on the screen, taps, then checks what actually changed.
The main part of the video is one take, in real time. No cuts, no speed-up.
1
45
It’s free and open source.
14 updates in its first week, most of them from bugs I hit using it myself, and I’m not slowing down.
If you build iPhone apps and don’t own a Mac, I’d love for you to try it 👇
github.com/bo5amis/tapwright
40
Mohammed Bin Khamis retweeted
حكايا المشموم 🌱.. تروي لنا السيدة ظبية السليطي، مستشارة ومحاضرة في التوثيق اللغوي والهوية والبيئة القطرية، جانب من تجربتها وشغفها بالمشموم كونه جزء جميل من الثقافة والتراث في البيت القطري.
وانتو .. بشنو يذكركم المشموم؟
شاركونا بحكاياتكم
Help. I can’t stop spinning up dumb things to work on for no other reason than to have fun building things and testing them.
@t3dotcodes @theo
1
1
299
Tapwright 1.6 is out: the Speed Update ⚡
An AI agent driving a real iPhone, measured against 1.0 on the same phone, app and test runner:
Tap: 18% faster Toggle: 30% faster Scroll: 20% faster Wrong-element misses: halved
Real time, no cuts 👇
reading the screen and typing did not get faster. iOS building its accessibility tree sets that pace.
Over Wi-Fi, gaps under 20% are noise.
Full table, setup and caveats: github.com/BO5AMIS/tapwright…
1
37
Free and open source. Drive your iPhone from Windows, macOS or Linux. No Mac needed.
pip install tapwright
github.com/BO5AMIS/tapwright
37











