Solutions
AI training data, and the one step we serve
This URL used to advertise GPU clusters. We do not own any, and the rest of what it promised was invented too. What is actually here is the first step of a data pipeline: fetching public documents at volume, with enough per-request detail that the corpus is accountable later.
What we do not offer
There is no compute here. No GPUs, no clusters, no managed training, no AutoML, no experiment tracking, no model hosting and no inference of any kind — the LLM product this site once advertised was removed from every customer surface because it could not serve a request. We also sell no datasets, maintain no corpus, and do not check whether you are licensed to train on what you fetch.
The unglamorous first step
Model work fails at the data step more often than at the training step, and the data step usually starts as an HTTP problem: a large list of URLs, fetched from addresses that do not all get rate-limited at the same moment, with enough detail per response to know what you actually received rather than what you hoped for.
That is a narrow, real thing to buy. It is also the only part of this problem we can honestly sell you, so it is the only part this page describes.
Provenance is the part that hurts later
A corpus you cannot account for is a corpus you cannot defend. Months after a training run, someone will ask where a document came from, whether it was fetched once or four times, and whether a retry storm quietly triplicated a section of the data. The answer has to have been recorded at collection time.
Idempotency-Key
A retry after a timeout is neither re-executed nor billed twice, and the replay is marked Idempotency-Replayed: true. Duplicates in a training corpus are not a billing problem, they are a weighting problem.
The target's own status
A 403 block page is well-formed HTML and will happily become a document if nothing checks. The response reports the status the target answered with, so you can drop them before they reach the corpus.
final_url
Redirects mean the document you received may not live at the URL you asked for. Record the URL that actually produced the body as the provenance, not the one you sent.
content_type and bytes
The cheapest filters you have. Both come back on every fetch, before you spend anything on parsing or tokenising.
Rights, robots, and what we do not know
We do not check whether you are permitted to train on what you fetch, and we could not if we wanted to. Egress nodes do not log target URLs or response bodies, so we hold no record of what you collected. That is a privacy property first, and it has an honest consequence: we cannot reconstruct your corpus for you and we cannot vouch for it.
Our terms forbid bypassing authentication, paywalls or access controls, and forbid scraping in violation of a site's terms where you have been put on notice of them. robots.txt, licence terms and applicable law are decisions you make before the request goes out, not ones we make after it comes back.
Retrieval at inference time is the same call
The other place this fits an AI system is retrieval rather than training: an agent or a RAG pipeline that needs the current version of a page instead of a model's recollection of it. Same endpoint, same envelope, one URL at a time, with the region chosen per call.
There is an MCP server, so an agent can call the fetch as a tool without you writing a client for it. Copy-paste snippets and a generated Postman collection sit alongside it in the dashboard. Python and JavaScript clients exist in our repository but are not published to any package index, so do not expect to install them by name.
What you build
The URL frontier, extraction and cleaning, deduplication, quality filtering, tokenisation, the store, and the training itself. All of it. We are one call inside the first of those, and being clear about that is more useful to you than a longer list would be.
Questions
- Do you run models or provide inference?
- No. There is no inference product, no model hosting and no compute of any kind. An LLM product was advertised here previously and could never serve a request; it was removed rather than left up.
- Do you sell a training dataset?
- No. We hold no corpus and no archive. Nodes do not store response bodies, so there is nothing here to sell even if we wanted to.
- Does JavaScript-rendered content come through?
- No. You get the HTML the server sent. We run no headless browser, so a site that assembles its text in the browser will yield a shell. For corpus work that is often fine, because the sources worth having usually serve their text.
- Can my agent call this directly?
- Yes. There is an MCP server for tool-calling agents, and the HTTP API itself is a single POST with an API key header, which most frameworks can call without a dedicated client.
- Does fetching through you give me the right to train on it?
- No. We are transport. Whatever rights exist are between you and the source of the content, and our terms make honouring them your obligation.
Get a key
Create an account and mint an API key in the dashboard. The full endpoint reference — request shapes, parameters and error codes — is published at https://api.roamingproxy.com/v2/docs.