YouTubeWhat they did
The author wanted to know what YouTube's web player downloads before playback, what it reports while a video plays, and how it picks a quality. He installed REA 6.8.0, an open-source reverse-engineering toolkit that runs as a CLI and as an MCP server, into Claude Code with its skill, and attached it to a separate Chrome profile over the DevTools port. He ran REA's capture, execution-tracing and bundle-analysis tools on two videos, throttled the connection to force quality changes, and checked every answer against his own DevTools network log and the player's stats panel. He also asked Claude Code the same questions in headless runs with the MCP server registered.
How it went
REA's capture matched his log request for request: about 81 requests in 28 seconds, 13.7 MB of JavaScript across 50 scripts, and the timing of the watch-time and quality reports, with a call stack for each. It did not find where the quality decision lives: static analysis timed out on the two scripts that matter even with 10 minutes each, and execution tracing returned only minified function names. Of three headless Claude Code runs, the first spent its 12 turns reading a 1.17 million character result, the second picked a tool that records only navigation, and the third, with the tool and parameters named, answered in seven turns and 50 seconds.
Worth knowing
Name the REA tool and exact origins in the prompt, and tell the agent to read large results from a file with jq; left to choose, it picked badly. REA only accepts exact origins, so wildcard video hosts are refused. Use a throwaway Chrome profile, since anyone who reaches the debugging port controls the browser.