Terms of use

Draft of 14 August 2026

These terms cover the SKYGEN engine: a research and outreach service that finds business contact data from public sources and, when a person approves it, sends email. By using it you agree to them. If you are using it for an organisation, you are agreeing on that organisation’s behalf.

It is free, and there is no payment

Everybody is on one free tier. There is no card, no invoice, no payment processor involved, and nothing to cancel — stopping means not using it. We have not agreed a price with anybody and nobody has paid us anything for this service.

If a paid tier is introduced later, it will be something you have to accept before it applies. Nothing you are using today starts charging by itself.

No service level, and no promise it stays up

There is no uptime commitment, no response-time commitment, no support commitment and no backup commitment. This is early software running on infrastructure we do not own. It can be slow, it can be down, it can lose a world, and features can be removed.

What we do commit to is the thing the product is actually about: when it does not know something, it says so. A failed lookup comes back as a failure, an unverified address comes back marked unverified, and a run that was rate-limited says it was rate-limited. That is a property we maintain deliberately, and reporting a failure as an empty result would be a defect, not a stylistic choice.

Your account

You can use the product without an account: a cookie says which browser is asking, and what you build is kept against it. Signing up claims the worlds the server already holds for your browser. Work an anonymous visitor built entirely in the browser is not migrated — it stays in that browser and is still there when you sign out. We would rather say that than let you find it. Keep your password to yourself — anybody who has it can drive your engine and read your lists, and we cannot tell the difference.

Accounts are for the person or organisation that created them. Do not share one across companies whose lists should not see each other.

The limits you are working inside

There are no per-account allowances any more: no cap on sends, agent runs, model tokens, worlds or documents, and nothing to upgrade to. What is left below is not an allowance — it is what one deployment can serve, and what keeps our crawler and our verifier from ever looking like an attack on somebody else’s server. These are today’s numbers and they change without notice:

  • 200 worlds per account.
  • 6000 reads, 3000 writes and 600 heavy operations per minute.
  • 25 messages per mailbox per day, with at least 90 seconds between two sends from the same mailbox.
  • 20 address-verification probes per domain per hour.
  • At least 0.4 seconds between two requests to the same website, and longer whenever the site asks for longer.

The last three are not quotas we would sell you out of, and removing the account allowances did not touch them: they exist so that neither our crawler nor our verifier ever looks like an attack on somebody else’s server, and no amount of demand changes them. Sending still waits for a person every time, at every autonomy level.

You are the controller of the list you build

This is the most important clause here. You decide which companies are researched, which people end up on a list, what the message says and when it goes out. Under data protection law that makes you the controller of that personal data and us your processor for it. The terms of that relationship are in the data processing addendum, which forms part of these terms.

In particular, you are responsible for having a lawful reason to contact each person, under the law of their country as well as yours. Several countries require prior consent for unsolicited commercial email even between businesses. The product does not check this and does not warn you. It will happily send where sending is not allowed, because it has no way to know.

Two things do not belong to you and cannot be assigned by these terms: what the engine works out about the world goes into a corpus shared across accounts, and the suppression list is kept for the person who asked to be on it. Both are explained in the privacy notice.

What you must not use it for

Breaking any of these is grounds for suspension. Most of them also produce a complaint, and the complaint lands on us.

  • Consumer marketing. Business addresses, at a business domain, about that business. Not private individuals as individuals, not personal mailboxes, not a list of people who are not at work.
  • Sending where consent is required and you do not have it. Your judgement, your responsibility, and see the clause above about the product not checking.
  • Writing to a compliance desk. The engine refuses to promote gdpr@, dpo@, abuse@, postmaster@ and their relatives to a send target. Do not route around that by hand.
  • Using verification as a directory attack. Probing a mail server for addresses you have no reason to think exist, or enumerating a domain, is a misuse of the feature and of somebody else’s server.
  • Misrepresenting anything. Who is sending, what is on offer, or where you got the address.
  • Targeting a person rather than a business. No harassment, no stalking, no tracing an individual, no building a file on somebody.
  • Anything illegal or hostile. Phishing, malware, fraud, and the categories that make an entire sending estate untouchable — crypto solicitation, pharmaceuticals, adult content, debt or loan offers to strangers.
  • Defeating the crawler’s manners. No proxies, no spoofed fingerprints, no ignoring a site that asked us to stop. If a site tells us to stop, that decision covers your account as well as ours.
  • Reselling the output as a contact database, or running the service as a backend for one.
  • Loading it beyond the published limits, or automating around them.

A send always needs a person

A set of capabilities is refused at execution unless a person approves — with one exception we would rather name than bury: a scheduled reply poll, which reads the mailboxes on a timer behind a shared secret and no person them — sending email, buying a domain, provisioning a mailbox, warming a domain, reading a reply mailbox, taking somebody off the do-not-contact list, and recording that a creator was contacted. The planner does not offer them and the executor refuses them, so an agent cannot decide by itself to write to anybody. An agent also cannot widen its own permissions — the autonomy level it runs at is never something it can change.

That set is computed, not listed. It is derived from the tool registry — everything that reaches money or mail, plus the handful that record something only a person can know — so a capability added later is inside the gate on the day it is added rather than on the day somebody remembers to amend a sentence here. A count in this paragraph would go stale; the derivation does not.

That is a property of the system, not a promise about your staff. Whoever presses the button is doing so as you.

Suspension and termination

We can suspend or end an account that breaks the rules above, or that is generating complaints, and we will tell you which rule and what we saw. Where the problem is fixable and the harm is not ongoing, we will ask first.

We act on complaints rather than on a bounce rate, because bounce rate is not measured yet — the only such number in the product today is a placeholder, and suspending an account on a number we cannot stand behind would be exactly the thing this product exists not to do.

You can stop at any time. Ask us and we will delete the account.

No warranty, and what the output is worth

The service is provided as it is, with no warranty of any kind — not of accuracy, not of fitness for a purpose, not of uninterrupted operation. In particular, read what a result claims before acting on it:

  • inferred means nobody has confirmed the address. It is a rendered guess with a source attached.
  • unknown means we did not find out. It does not mean invalid, and treating it as invalid deletes real people from lists.
  • A domain that accepts everything has proved nothing by accepting one of ours, and results from it are reported as unproven.
  • A name from a search index carries no leaving date. Some of those people have moved on.

Sending is what damages a domain’s reputation, and the domain is usually yours. The caps and the verification exist to reduce that risk; they do not remove it, and we do not warrant deliverability, replies or any commercial outcome.

Liability

Nothing has been paid for this service, so there is no refund and no fee to compute a cap from. To the fullest extent the law allows, we are not liable for lost profit, lost pipeline, lost data, a damaged sending domain, a blocked address, or any indirect or consequential loss. Where the law does not permit an exclusion — death or personal injury caused by negligence, fraud, and anything else a court will not let us disclaim — our liability is whatever the law requires and no more.

If your use of the service causes a claim against us — a complaint from a person you wrote to, a regulator, or a site you crawled through us in breach of these terms — that claim is yours to answer.

Your data, and getting it out

A world is readable through the same API the interface uses, so what you built can be fetched as JSON today. There is no packaged export yet — no button, no archive of everything an account holds. Ask and we will send you what there is.

What happens when it ends

On termination, by you or by us: the account, its worlds and its memory are deleted. Ask, and we do it.

Worlds are meant to be deleted automatically after thirty days without activity, whether or not the account is open. That collection runs on a timer inside a long-running server process, and nothing schedules it on the edge runtime this deploys to — so today it is something to ask for rather than something that happens, and you should read the thirty days as an intention rather than a guarantee. It is on the list below.

Two things deliberately survive, and one of them is not about you:

  • The suppression list. Somebody asked not to be written to, and that record has to outlive your account or the request stops meaning anything. An opt-out taken through the link in a message, or by a reply, is written against your account and refuses every campaign and every mailbox you have. Taking one off again is possible, it is recorded with a reason and who did it, and it is not something the engine may do on its own. The privacy notice tells the people it affects the same thing.
  • What the engine learned about companies. The shared corpus is not held on your instruction and is not yours to withdraw. See the privacy notice and the addendum, which both say this in as many words.

Changes

These terms will change. The date at the top changes with them, and a change that materially reduces what you can do is announced to customers before it takes effect. Continuing to use the service after that is how you accept it.

What is still open in this draft

  • The legal entity offering the service, its address, and where notices are sent.
  • Governing law and the courts that hear a dispute.
  • Whether an indemnity is stated as a clause or left to the general law.
  • How much notice a change of terms gets, in days.
  • A packaged export, so this page can stop saying there is not one.
  • A scheduler at the edge, so the thirty-day world deletion above and the mailbox read that honours a “stop emailing me” reply both happen without somebody remembering.