← zrxiv

Terms of use

These terms cover zrxiv.dev: the website, its JSON API (?format=json on search and record pages, /stats and /blob) and its public MCP endpoint at /mcp.

Who runs this

zrxiv is operated by Pogoapp LLC, a limited liability company organised in Delaware, United States, as part of the Epistack project. Write to contact@zrxiv.dev.

What you may do

Read, search and link to anything here. Machine access through the JSON API and the MCP endpoint is welcome, including by agents, and needs no key and no account. Where you can, link to a work by its DOI (/doi/…) or arXiv id (/arxiv/…): those addresses do not move. They work for a work we do not hold yet, too: its page then shows what the public indexes say about it, and asks us to add it.

What you may not do

• Bulk-download or scrape full text from this site. Where a work's licence allows reuse, take it from its publisher or repository; if you need a corpus, several of the upstream sources listed on our home page offer bulk access built for that.
• Work around a withholding. A withheld page, an HTTP 451 or a refusal from zrxiv_fulltext means we have decided not to serve those bytes.
• Try to overload the service, or probe it for security weaknesses without asking us first.
• Send requests to add a work that you do not mean: each one is real work for us.

Requests to add a work

A request to add a work (POST /doi/…, POST /arxiv/…, the MCP tool zrxiv_request, or opening the page of a work we do not hold yet) may fetch and store an openly licensed copy of it, from the lawful location that states the licence (CC BY, CC BY-SA, CC0 or public domain); nothing else is fetched. Where no such copy exists, we add the work's citation: its metadata and a link. Asking does not make a copy open, and does not entitle you to one: the licence at the location we fetch from decides, and this site's tier map and takedown list still apply to what it serves.

Rate limits

Requests to add a work (POST /doi/… and POST /arxiv/…) are limited per client, counted by IP address: 20 a minute, in bursts of up to 20. Past that, the answer is HTTP 429; try again shortly. Requests are also paused, with HTTP 503, while the queue or the store's disk is full. Opening the page of a work we do not hold yet counts as a request to add it; past the limit the page still opens, and says it did not ask.

Through the MCP endpoint, zrxiv_request is held to the same limits and to a slower ceiling of its own: 10 new requests per client per hour, counted the same way.

When zrxiv's own reading assistant asks for a work on a visitor's behalf, the request is counted per conversation instead: 3 new requests per conversation per hour, and at most 60 per hour and 300 per day across all of its conversations together.

Reading (pages, the JSON API, /blob and the MCP endpoint) has no per-client limit in this service, with one exception: the MCP tools that look a work up in outside indexes (lit_search, zrxiv_locate and zrxiv_reading_context), and the page of a work we do not hold yet (/doi/…, /arxiv/…), which looks it up in the same indexes, may do so at most 30 times per client per 10 minutes between them, counted by IP address like requests to add a work, so that one client cannot use up the pace those indexes allow us; past that, the tool answers with an error, and the page with a Too Many Requests status, asking you to try again in a few minutes. All traffic reaches us through Cloudflare, which applies its own protections and may challenge or block traffic it judges abusive; some programming libraries' default User-Agents are challenged, so send a real one from scripts. We may add limits if bulk extraction makes them necessary.

When zrxiv's own reading assistant looks a work up in outside indexes for a visitor (the MCP tools lit_search, zrxiv_locate and zrxiv_reading_context), the lookups are counted per conversation instead: each conversation may make at most 20 such lookups per 10 minutes, so that one conversation cannot use up the pace those indexes allow us.

The licence of each work

Each work's own licence governs its reuse, not these terms, and we cannot grant rights we do not hold.

The licence gate is on here: a served work's page shows the licence on the copy we hold in its Licence row, and the JSON API and MCP endpoint report every work's tier and licence (the tier and licence fields). Where no licence is shown, none is granted.

What we serve and why is set out in the policy.

No warranty

The service is provided as it is, without warranty of any kind. Records, extracted text and chases can be wrong or incomplete, and a mirrored copy can differ from the version of record: check anything that matters against the source, at its DOI. We may change, pause or withdraw the service, or any record in it, at any time.

The reading assistant

The chat window on the home page and on each paper's page is a reading assistant: a language model that searches this site and public indexes of the literature and answers questions about papers. It is a program, not a person, and it can be wrong: it can misread a paper, confuse two works or state something no source says. It is instructed to cite where each claim comes from; check anything that matters against the paper itself.

It does not give medical, legal, financial or other professional advice, and nothing it says should be taken as such. It reports what papers say and how sure they are; it does not decide who is right. The assistant may read pages on the open web to answer you; what it reports from them is a summary of someone else's page, with its address, and not advice or a checked fact.

Use it to read, not to copy. Do not use it to extract the text of works this site does not serve, or to get around its limits. Before a reply is shown, it is checked for how much of any one work it quotes, and a reply that quotes too much is not shown.

Each conversation can send up to 30 messages an hour. Past that, the chat window says the limit has been reached, until the hour has passed. In a shared conversation the limit is each person's own.

A conversation on the research page can be shared with a link. What you ask in a shared conversation is shown to everyone who has the link, and so are the replies. Do not share a link with anyone you would not show the conversation to. We may remove a shared conversation at any time. A link shared with invited people only is the responsibility of the person who shared it: whoever they invite sees the whole conversation.

Research artifacts are an alpha feature, not a production service, and are available only to approved signed-in accounts. They are private to their conversation's owner: an artifact address is not a public or secret sharing link. Only plain text, Markdown source and CSV are supported, not HTML or executable pages. Conversations with artifacts cannot be shared. Check generated content before relying on it; CSV downloads neutralise formula-like cells.

The assistant has a daily budget. When it is used up, the chat window says: "The assistant is resting until tomorrow. Everything on zrxiv is still here." It opens again at midnight UTC. If the assistant's service is down, the chat window says it cannot reach it, or does not appear at all. The rest of this site does not depend on it.

Transcripts of conversations are kept on our servers and deleted 30 days after a conversation's last message. What a conversation leaves behind is set out on the privacy page.

The assistant is run by the Epistack project, like the rest of this site (see "Who runs this" above). We may change, limit, pause or withdraw it at any time.

Changes

We will revise these terms and the privacy page before any account feature ships on this site, and whenever the assistant's limits or what it keeps change.

Notices about content we serve: takedown. Everything else: contact@zrxiv.dev.