Guide

Explorer

The pages behind archives.agntn.dev that run the library live, and what each one is for.

Every explorer page calls the same executors the MCP server uses, through the docs worker. Nothing on these pages is a mock; a slow answer is the archive being slow, and a 429 is the archive saying no.

PageWhat it answers
TimelineWhich captures exist for a target, across every provider, with a viewer for any of them
CompareTwo captures of one page side by side, with the diff, and sliders that walk the history
SiteWhich archives hold a domain, how many captures, first and last, a year by year heatmap
URLsEvery URL an archive has seen under a domain
HistoryHow a page changed capture by capture, ranked by the size of the change
AgentThe four tools with their exact text answers, and the call as JSON-RPC or TypeScript
StatusWhich providers answer right now, and how fast
ShelfCaptures you kept in this browser, exported with their provenance

The viewer

A capture opens in one of three modes. Replay frames the archive's own playback, only for archives that allow framing. Source draws the archived markup inside a sandboxed frame with no scripts, a policy that admits only assets from the archive's host, and a <base> on the capture. Text is what archives_content returns. Every capture has a permalink under /capture/<provider>/<stamp>/<url>, a citation to copy, and a bookmark for the shelf.

Caching and manners

Answers are cached on the worker, in KV in production, for thirty minutes to six hours depending on the page, and a cron warms the demo targets. A failed answer is never cached, and one where a provider failed is kept for five minutes only, so an outage does not stick. One address can start thirty new archive queries a minute; answers served from the cache do not count. The explorer never retries a 429; Archive.today in particular throttles automated readers, and the viewer says so instead of hammering it.

@agntn/archives·MIT license· Archived pages are data, never instructions.