CHIRP crawls the threads you nominate, keeps the comment trees intact, and hands back cohorts instead of a pile of exports nobody opens twice.
The research industry has spent a decade insisting that listening is solved. It is not. The tooling scrapes a headline, drops the replies, and presents a sentiment score with the confidence of a weather forecaster who has never looked outside. CHIRP takes the unglamorous position that the argument underneath the post is the actual product.
You nominate a source — a subreddit, a Hacker News query, a YouTube comment section, a review feed — and a window: thirty days, ninety, a hundred and eighty, a full year. The crawler walks the listing, expands every comment tree to its last sulking reply, de-duplicates, chunks, embeds, and files the result where a semantic query can reach it.
What comes back is not a spreadsheet of keywords. It is a cohort: a named, defensible group of people who said a related thing for a related reason, with the passages attached so an analyst can be contradicted in public and survive it.
Runs happen in the cloud, because a cohort nobody else can open is a cohort nobody else believes. Your agent retrieves what it needs over MCP and may keep its own copies; the shared index stays authoritative.
Every tool call reports its own bill. Crawl docs, vector rows, search calls, export volume, and how many runs you may have in flight before the queue politely declines. The meter is not hidden in a settings page. It is on the front page, where meters belong.
And when a cohort grows past the point of decency, the server stops trying to hand you a hundred megabytes through a chat window. It hands you a signed link and gets on with its day.