AI & Agent

A five-minute walkthrough of why agents should use the Chrome you're already signed into, how chrome-use connects to it, how it leaves your tabs alone, and what happens at a captcha.

motion-use is an open-source video tool. Your coding agent writes exact-time drawing code, renders it locally, then decodes the finished MP4 and checks it. With a one-minute film where every clip shown was made with it.

An open-source command-line tool that lets agents like Claude Code and Codex operate a Switch running Atmosphère and sys-botbase over a local network: OCR for reading screen text, tapping by text, waiting for screens to appear, buttons and sticks, touch, and screenshots.

A one-minute iphone-use video where every phone shot is a real recording. The AI reads the screen, taps by name, says when a step didn't land, and replays finished tasks without spending tokens.

With a single request, Opus 5.5 wrote the director’s notes, three.js scenes, a music synthesizer, and subtitles in Claude Code; reviewed extracted frames itself to fix bugs; then used the open-source motion-use to render a 102-second, 6,120-frame finished video locally.

I run Claude Code, Codex and Pi side by side, and they can't see each other. ocs gives every AI coding session a shared channel and actually wakes the target session — on one machine or across the computers on your LAN. How it works, and the "it said OK but the message was gone" bugs I had to design around.

When multiple agents develop in parallel, Rust repeatedly recompiling dependencies is often the biggest cost. Limit concurrency across two axes, share caches, isolate target directories, then pin the rules in AGENTS.md so your computer can handle three active lines of work at once.

AgentParty is the networked version of cross-session: Claude Code’s cross-session only works between your own sessions, while AgentParty extends it across different people, teams, and companies, letting my AI talk and debug directly with your AI and eliminating the human translation layer in between.

AgentParty is a very thin bus that lets coding agents from different machines and different companies — along with the people behind them — stay in the same channel to @ each other, stand by, and hand things off. This article explains it along three lines: what it is and what problem it solves; what scenarios it can be used in; and how its underlying architecture is built.

A walkthrough of the full principles behind video generation using the real Wan2.2 code: flow matching mathematical derivation, VAE spatiotemporal compression, DiT and 3D RoPE, CFG, MoE dual experts, condition injection, and sequence parallelism, with concrete numbers calculated for every concept and original code excerpts included.

From the five-task matrix and layered architecture to DiT/VAE/T5/distributed execution/schedulers, this article breaks down, layer by layer, the real pitfalls in the Wan2.2 source code where the documentation and implementation diverge.

Engineering notes from three months in production: typing rhythm, model selection, layered prompts, guaranteed-reply fallbacks, plus three reversals that took cache hit rate from 61% to 91%. Building an AI chat companion that does not break immersion is not about the model; it is about engineering details.

chrome-use’s selling point is that it lets agents use your one real browser. But that brings a problem that tools which open a new browser do not have: several agents plus you may all be acting inside the same Chrome at the same time. This article explains how it isolates work by “who opened the tab” (colored tab groups + a dedicated daemon + only touching its own pages), and one trade-off that can trip you up if left unsaid: this isolation only works through the extension connection path; a bare --cdp port is not isolated.

I tried three ways to fetch data: stuffing a full HTML page into context, doing one find/get round trip per field, and hand-writing eval scripts. All felt awkward. What remains is chrome-use’s site and extract, based on the same judgment: data extraction should be a single declarative evaluation, not a chain of imperative round trips. This article explains how to use both from real output, how to write an extract schema correctly (I got it wrong at first), and where it cannot or should not be used.

The first time I connected chrome-use to the Slack desktop app and got a snapshot full of named buttons, I was taken aback. Later it clicked: the Slack desktop app is a web page to begin with. An Electron app is the Chromium rendering engine wrapped in a native shell, with a clearly structured DOM inside. This article explains why that turns the traditionally messy work of “desktop automation” into operating a web page, and one trade-off you need to know compared with browser scenarios.

Anyone doing browser automation will eventually fall into the same trap: desperately trying to disguise a fake browser as a real human, only to be exposed layer by layer. To understand this, you first need to know that detectors like CreepJS do not directly decide whether you are a bot. They collect dozens of signals from your browser and cross-check them. Every mismatch counts as a lie. This article starts with a real fingerprint I tested, breaks down the hardest trace to hide, which lives at the protocol layer rather than the JS layer, and explains why patching is a negative-sum game.

I compare chrome-use’s connection to an existing Chrome session with the new and persistent contexts used by Playwright, Puppeteer, and browser-use, explaining session state, tab ownership, connection requirements, and the limits of fingerprint testing.

During development and debugging, want to mock an unfinished API, change a request parameter, or modify a real response? chrome-use network route intercepts directly in the Chrome instance you are driving, with four uses: mocking responses, changing requests, modifying real responses, and blocking. This article explains why it is the lightest path—built on CDP Fetch, with zero JS injection, no added extension permissions, and no Chrome Web Store re-review—and includes real-device tests.

chrome-use test turns manual browser checks into repeatable YAML suites that can run in CI, effectively writing unit tests for the frontend. This article explains what it is, why its implementation can be so thin, and how to use it with cookie-use to switch back and forth between multiple accounts.

I wanted an AI agent to log into accounts by itself. The username-and-password path worked with bitwarden-use, but the OAuth path got stuck for most of a day. This explains how the three windows behind “Sign in with Google” talk to each other, why automation breaks as soon as it touches them, the three cuts chrome-use made, and the final fully automatic login to X.

We spent the whole night chasing hooks, proxies, packet captures, and session logs, only to find that the scariest piece of “injection evidence” was not tool output, but a story Claude had told itself.

Stop launching automated browsers from a blank profile, logging in again, and running into CAPTCHAs. chrome-use lets any AI agent directly drive your real Chrome, where you’re already logged in to everything: CreepJS rates it 0% bot, and with structured snapshots plus an @ref interface, reading a page costs only 200–400 tokens without burning money on screenshots.

People are sharing claims that Codex secretly writes logs like crazy and burns through SSDs. Local test: the file stays stable at 126 MB thanks to rolling cleanup, but over 10 days it inserted 8.45 million rows, and the writes came from the background desktop app-server. The original post is basically directionally correct; only the “everyone will definitely burn through an SSD in a year” part is exaggerated. Includes root cause (source code + GitHub issue) and cleanup options.

Let agents like Claude Code and Cursor generate images along the way when writing docs or READMEs, without an OPENAI_API_KEY, without paying extra, and using the ChatGPT subscription you already have. This article dissects the most interesting implementation path behind it: the web backend. Why you can’t just POST directly, why the real wall among the three layers of anti-scraping is the single-use Turnstile token, and the complete flow for driving your already-logged-in Chrome to generate images.

When an AI agent needs an image while writing, the traditional path either requires an API key and money, or a human has to go to ChatGPT, generate the image, and paste it back in—the agent can only get stuck waiting. chatgpt-imagegen lets the agent generate images on its own using your existing ChatGPT subscription: no API key required, no Codex quota consumed by default, and image-to-image generation supported. This article explains its original intent, how its two backends work, and why it was designed for agents.

When connecting an AI agent to a browser, the hardest part is not opening pages—it is the forms hidden inside cross-origin iframes. The agent can read them and fill them in, but it just cannot click “Save.” A look at how this tough problem was solved.

Open-source real iPhone automation with a custom XCTest runner, CLI, HTTP API and MCP: read the screen, tap, type, remotely control and replay workflows.

Claude Code released dynamic workflows yesterday (2026-05-28), and it is still a research preview. This article does not discuss source code; it only explains what it actually changes, when it is worth using, when it is purely a waste of tokens—and the essential differences between it and subagents / skills.

How the status bar works out whether Claude Code’s prompt cache is still warm and how many minutes are left, when the default 5-minute TTL is right, why 1-hour cache users see “cold” 55 minutes too early, and where the countdown can lie.

A one-page visual overview: how a ChatGPT subscription is turned by chatgpt-imagegen into a command-line image generation tool—OAuth token retrieval, Responses API SSE streaming, capability boundary matrix, and scenarios where you should not use it, all explained at a glance.

Based on the Claude Code 2.1.139 headless source code, this breaks down how the /goal command drives the Agent to keep working until a condition is met through a session-level Stop hook.

What actually happens when you run claude --bg, /bg or claude agents: one on-disk job model, a shared daemon (with bg-spare and bg-pty-host helpers) and the FleetView TUI, traced through the Claude Code 2.1.139 source.

A Slidev-style Chinese presentation translated into English, explaining Agentic Coding, Skills and MCP, AGENTS.md and CLAUDE.md, sandboxing and permission controls, and how a team can truly operationalize AI programming with Plans, a YApi Skill, and docs-sync.

A systematic overview of the core questions around Agentic Coding, covering the main forms of today’s tools, the boundary between Skills and MCP, how AGENTS.md and CLAUDE.md work, sandboxing and permission control, and how to evaluate AI programming with real tasks and benchmarks rather than demos alone.

This article goes beyond the broad “control plane / execution plane” framing and instead walks through the current implementation piece by piece—covering the bridging layer, IPC, configuration fields, the WeChat flow, rollback mechanisms, and the evidence index—to explain how QClaw organizes OpenClaw into a deliverable desktop runtime.

Based on the official release notes, this article summarizes the key changes in OpenClaw v2026.3.8: backup commands arriving in the main workflow, macOS remote gateway onboarding, Talk silence timeout, Brave search integration, ACP receipts, and multi-platform routing fixes.