Anna's Archive published a page on February 18, 2026 with one specific addressee: LLM crawlers. The site holds 64,416,225 books and 95,689,473 papers, has been served behind CAPTCHAs designed to deter bulk scraping, and has now written a polite, machine-readable note asking model trainers to please use a different door. The page is at annas-archive.gl/blog/llms-txt.html, with a permanent landing copy at annas-archive.gl/llm.
The page is two things stacked. The surface artifact is an llms.txt file — a young convention that mirrors robots.txt but addresses language models instead of search crawlers. The substance, written into that file, is a four-step routing guide off the human-facing site and onto bulk endpoints that the project would rather see crawlers use. The post resurfaced on Hacker News this week with 750 upvotes and 413 comments, three months after publication.
What the page actually documents
The llms.txt file names four ways to ingest the collection without round-tripping through the public search UI:
- The GitLab repository at software.annas-archive.gl carries every HTML page and all the site's own code.
- The Torrents page exposes metadata and full files; the project flags
aa_derived_mirror_metadataas the entry point for trainers who want the catalog. - The Torrents JSON API at
/dyn/torrents.jsonlets a crawler enumerate the torrent set programmatically rather than scraping the listing page. - The donation-tier API at
/faq#apireturns individual files after the requester makes a donation. A separate enterprise tier at/llmadvertises SFTP access to the full corpus for donations in the "tens of thousands USD" range — refundable in trade for OCR, deduplication, or text and metadata extraction work.
The page is explicit about the ask: "Our website has CAPTCHAs to prevent machines from overloading our resources, but all our data can be downloaded in bulk." Translation: do not burn budget breaking the CAPTCHAs; use the door we built for you. The Monero address at the bottom of the file is the unfunded version of the same pitch.
Why a shadow library writes a cooperation overture
Anna's Archive is not in a position to enforce anything against an LLM trainer that decides to scrape the site anyway. The llms.txt convention has no teeth — robots.txt doesn't either, and the established crawler-compliance record suggests the major labs read it when it suits them and ignore it when it doesn't. So why publish this?
The structural bet has two parts. First, the economics. The trainers who would honor the request are also the ones whose compute budgets make CAPTCHA-bypass economically painful at the scale of 64 million books. Pointing them at torrents and SFTP saves both sides money — and the project's pitch on the /llm page is candid that the dollars saved on bypass infrastructure could be redirected as donations. Second, the optics. By documenting the routes explicitly, the project moves itself from "data has to be taken" to "data is being offered" — and offers an enterprise-paid channel that some trainers may eventually want as a defense against discovery requests asking how their training set was acquired. Neither claim is the legal cover the trainers would actually need, but the framing is the cheapest part of the trade.
The broader signal here is about llms.txt itself. The convention was introduced for documentation sites that wanted to publish a curated, LLM-friendly version of their content for crawler consumption. A shadow library adopting it for the opposite use case — bulk-data licensing optics — stretches the convention into territory its authors probably did not have in mind. Expect more of this kind of usage as robots.txt-style files continue to gel into the de facto contract between sites and the trainers reading them.
For a builder running retrieval over books or papers, the practical takeaway is narrow. If the project being built was already going to pull from Anna's Archive, the documented bulk endpoints are faster, cheaper, and less likely to break than scraping the public site. The legal exposure does not move because the endpoints are now documented — the underlying corpus is the same set of works whose copyright status got the project DMCA notices in the first place, and consulting in-house counsel before training on any of it remains the operative advice. What changed on Tuesday is that the routing instructions are now in plain text.
