sderosiaux/chrome-agent
What It Is
sderosiaux/chrome-agent is an independent Rust command-line project for agent-oriented browser automation through Chrome DevTools Protocol (CDP). Its public repository describes compact accessibility-tree snapshots, stable backend node identifiers, browser actions, inspection, and content extraction as parts of its interface.
How To Use It
The project's documented quick path is npx chrome-agent --help for a no-install run, or npm install -g chrome-agent for the prebuilt CLI. Its default loop is inspect, act, inspect: navigate with --inspect, use the returned UID, then request another inspection after an action.
# Start from the built-in CLI guide, then navigate and receive an a11y snapshot.
npx chrome-agent --help
chrome-agent goto https://example.com --inspect
# Act on a returned UID and immediately inspect the next state.
chrome-agent click n12 --inspect
# Extract content without writing a selector.
chrome-agent read
chrome-agent extract
# Obtain structured output when another program or agent consumes it.
chrome-agent --json inspect --urls --filter linkFor a real application, use inspect to get current UIDs, fill or select by UID, wait for an observable condition instead of sleeping, then inspect again. UIDs remain useful while their underlying DOM node exists, but a navigation or re-render can invalidate them.
Effective Use Cases
| Situation | Useful approach | Why this tool fits |
|---|---|---|
| LLM-driven browsing where context size matters | Use goto ... --inspect, then operate on compact a11y UIDs. | The project is designed to reduce page-state tokens and avoid hand-authored selectors for ordinary interactions. |
| Repeated list, feed, card, or table extraction | Start with extract; for React-style pages, use extract --a11y --scroll --limit N. | The CLI provides built-in repeated-content extraction before requiring site-specific scripts. |
| Readable article capture | Use read, then scope or cap output with text when needed. | Its Readability-based command is a concise alternative to inspecting an entire application DOM. |
| Fast interaction loops | Add --inspect to actions; use batch or pipe for a sequence of commands. | It combines state observation with actions and can avoid repeated process startup. |
| UI and API debugging | Use console --level error, network --filter, and wait network-idle. | The CLI exposes browser-visible errors, requests, and an observable readiness condition. |
Operating Tricks And Guardrails
- Start with
inspect --urls --filter linkwhen choosing among links: visible labels alone may not identify their destination. - Use
diffafter an action when a full snapshot would be noisy; use--max-chars,--offset, and--truncateto bound context returned to an agent. - Prefer
wait network-idleor a targetedwaitcondition to a fixed delay on SPA and XHR-heavy applications. - Scope screenshots to a UID or selector and use JPEG quality/width controls when visual evidence is needed without filling the agent context with a large image.
- Use named
--browserand--pagevalues to separate concurrent tasks; do not assume the default browser/session is appropriate for unrelated work. - Keep logged-in and personal data workflows authorized and minimal. This project documents cookie-copy and anti-detection features, but their availability does not authorize access, bypass controls, or bulk collection.
Why It Appears In This Wiki
This project was found accidentally during third-party validation of the similarly named chrome agent associated with Corey Gallon's scheduled World's Fair session. The initial repository match was wrong: this Rust project is not Corey Gallon's work. It is retained as a clearly labeled alternative because the name collision and technical overlap are useful comparison context.
Comparison Context
The scheduled chrome agent is Corey's Python CDP CLI at captivus/chrome-agent. This project is a separate Rust implementation at sderosiaux/chrome-agent; neither repository establishes facts about the other's author, release line, or World's Fair session. The implementation delta records the distinction.
Source Boundary
No official World's Fair schedule, speaker record, event recording, transcript, slide, or official AI Engineer source connects this repository to Corey Gallon or to the 2026 event. This page is supporting external comparison context from the project's public GitHub repository only, not first-class AIE evidence.