Leo

I build and test AI-agent tools for real Chrome, iPhones & everyday work. Sharing demos, failure cases and fixes. Full-stack engineer · ex-Qunar/Zhihu.

[Leo @leeguooooo]

2 minutes on iphone-use: an AI agent driving a real iPhone. Screen as text, taps that say when they missed, screens that block capture, tap + settle in 4.2s, and flows that replay with no model.

Open source: https://t.co/YYgCaFILmz

[Leo @leeguooooo]

@iamhenryyuan https://t.co/YYgCaFILmz

为什么要用硬件， 不是有现成的方案么

[Leo @leeguooooo]

How the AI did it — open-source *-use tools for Claude Code:
chrome-use (browser + captcha), mail-use (verification email), profile-use (personal details), bitwarden-use (password, card, 2FA). Video made with motion-use.
https://t.co/tT3Ww6VPen

[Leo @leeguooooo]

Free 2-core, 12 GB cloud server from Oracle — Always Free, $0/mo.

8-step walkthrough: sign-up, region, card check, launching an Ampere A1 VM, and what "Out of host capacity" means.

Fun part: an AI did the whole sign-up. I only answered 4 multiple-choice questions.

[Leo @leeguooooo]

@sweetbeetcoin @zhongerxin 我这就去优化第二个速度

[Leo @leeguooooo]

Part 2/2: how we built PR/deploy notifications on it, and the idea you can reuse.

[Leo @leeguooooo]

Your agent opened the PR. The deploy broke. Why are you still copying the error over?

We built PR/deploy notifications on open-cross-session: results go back to the session that opened the PR. 3-min walkthrough:

https://t.co/0ZYB6L4Jsi
https://t.co/m94VzOwDkG

[Leo @leeguooooo]

Part 1/2: what ocs is and how it wakes another session.

[Leo @leeguooooo]

Once a task works, save it as a flow. It replays in one call: no model, no tokens.

"How many steps did I walk each day this week?" comes back in about 20 seconds.

[Leo @leeguooooo]

My coding agent can now use my iPhone.

iphone-use lets Claude Code, Codex or any MCP client read the screen as text, then tap, type and swipe in any app, even Health, banks and chat apps.

USB or Wi-Fi, iOS 15+, open source, self-hosted.

https://t.co/YYgCaFILmz

[Leo @leeguooooo]

New since launch:
• iOS 15/16 phones over Wi-Fi, no cable, no iOS update
• live phone screen in the agent's side panel (MCP Apps)
• Codex plugin
• reads a list across pages and says when it hit the end
• refuses to tap through a system alert instead of pretending it worked

[Leo @leeguooooo]

Full 5-min walkthrough (separate agent window, 70+ site adapters, approval prompts, captcha handoff): https://t.co/VqhvU9yIye

Open source: https://t.co/H6S3pkhZhd

[Leo @leeguooooo]

Ask an agent to open a page and it usually gets a fresh browser: no cookies, no login, flagged as a bot.

chrome-use lets any agent drive the Chrome you're already signed into. No debugging port, no token, nothing leaves your machine. Works with Claude Code, Codex, Cursor.

[Leo @leeguooooo]

How it was made: Claude Opus 5.5 wrote this film as code with the motion-use skill; the phone shots are real recordings. The skill is open source: https://t.co/RheE1I2Jol

[Leo @leeguooooo]

iphone-use lets Claude Code and Codex use your iPhone the way you do. No app APIs: it reads the screen and taps buttons by name.

Every phone shot in this video is a real recording.

Open source: https://t.co/YYgCaFILmz

[Leo @leeguooooo]

How it was made: Claude Opus 5.5 directed and coded this film with the motion-use skill, then rendered and checked it locally. The skill is open source: https://t.co/RheE1I2Jol

[Leo @leeguooooo]

motion-use lets your coding agent make videos: it writes the film as code, renders it on your machine, then checks the MP4 itself (motion, loudness, misread narration, text off the edge).

Every film in this video was made with it.

Open source: https://t.co/RheE1I2Jol

[Leo @leeguooooo]

No frame of this 3D promo was made by hand.

Claude Opus 5.5 in Claude Code wrote 1,250 lines: the three.js scene, a synth for the score, subtitles. It reviewed its own frames and fixed bugs. Open-source motion-use rendered 6,120 frames.

https://t.co/RheE1I2Jol

[Leo @leeguooooo]

这支 102 秒的 3D 宣传片，没有一帧是手做的。

Claude Opus 5.5 在 Claude Code 里写了 1250 行代码：three.js 场景、配乐合成器、字幕，自己抽帧审片、修 bug。开源的 motion-use 在本地渲染出 6120 帧。

https://t.co/RheE1I2Jol

[Leo @leeguooooo]

Creating the account and adding the card both waited for my OK. Personal data in the video is masked; it's a reconstruction of the real run.

The *-use family (Claude Code plugins + CLIs): https://t.co/AUy8rmPSdJ

[Leo @leeguooooo]

I build *-use tools for #AIAgents. They just moved my paper water bill to card autopay from one line to #ClaudeCode:

chrome-use: the site
iphone-use: customer no.
profile-use + bitwarden-use: form, password
mail-use: email code
message-use: SMS (missed; iphone-use got it)