Elafi

What Elafi does in a crisis

Elafi is a wellbeing companion. It is not a crisis service, and it cannot send anyone to help you. This page describes exactly what it does when it notices someone may be in danger — including the parts that fall short.

If you need help right now

Built into the app, not generated by the AI. Last verified 2026-08-10.

How it notices

It runs on your device

The check that raises the banner is plain pattern matching in your browser — not a model, and not a request to our servers. Nothing is sent anywhere to make the resources appear, and what you wrote is never recorded because of it.

It reads more than your last message

In the main conversation it looks across your recent turns — up to 6 — because distress is often built up over a few messages rather than declared in one. Two different signals in two different turns are enough. In the couples space it deliberately reads only the turn being sent, so that one partner's earlier messages are not read as the other's — with one gap we would rather name: when both partners write before a reply arrives, their turns travel together, and on our side the check cannot always tell which words belong to whom.

It is checked twice

Our server independently re-runs the check on what it receives and combines it with your device's answer — neither side can switch the other off. What the server can change is the reply: a more capable model, and an instruction to name help. It cannot raise the banner on your screen, so if your app is out of date or broken, the resources may not appear even though the reply still treats the moment as serious.

The banner's numbers are not written by the AI

Every number and link in the banner is built into the app, versioned, and dated, so it cannot be invented or misremembered. Elafi's replies may also name 988 or findahelpline — those are written by the model, and the banner is the copy to trust. This list was last verified 2026-08-10.

What happens then

  • The resources appear, and they stay. A panel with the lines above opens at the top of the conversation. It cannot be dismissed. Once the moment has passed it shrinks to a small “Support lines” pill and rides out the rest of the day — in the main conversation it never simply vanishes. Two limits on that: it is rebuilt from the messages in your thread, so a banner raised by something you typed and then decided not to send lasts only until you reload, and the journal and the couples space are weaker throughout. See below.
  • If it looks immediate, emergency services come first. Wording that suggests something is in progress or planned raises a second tier: the panel cannot be collapsed at all, and your local emergency number is promoted above everything else. That number is chosen from your browser’s own language settings — there is no location lookup and no permission prompt — and if it can’t be determined confidently, the options are shown as plain text rather than dialled for you.
  • The conversation is never cut off. The chat is not blocked, the input is not disabled, and Elafi does not wind down or hand you off. Leaving is the one thing that would genuinely hurt.
  • The turn is escalated to a stronger model before any cost consideration, and the model is told to stay with you rather than recite — or, when the banner could not be raised on your screen, to name the numbers itself. It is also told, without exception, never to provide methods, means, or dosages. Two honest caveats: on a very heavy day the top tier can be exhausted and a crisis turn is answered by the next model down, and past the daily limit it is answered at that lower tier by design — and past a further, finite allowance beyond that limit, not answered at all until the next day. Not always the largest model available, then, and not without end.
  • What you wrote is never recorded because of it. The assessment itself is never stored, never synced, and never written into your notes — it is recomputed from your own words each time. Two counters do move on our side: one when the server notices something your device did not, and one when a conversation goes past the daily limit. They record that a check fired and nothing about what you said, and they carry no account identifier.

What Elafi cannot do

Elafi accounts are anonymous by construction: no email, no name, and everything you write is encrypted with a key we do not hold. That is a deliberate design choice, and it has a cost we will not paper over — we cannot identify you, contact you, or send anyone to you. There is no welfare check, no escalation to a human on our side, and no one reading your conversations who might step in. Nobody at Elafi can see that you are struggling.

We could only change that by building a way to de-anonymise you, which would break the promise the whole product rests on. So we have not, and the honest consequence is that in an emergency the people who can actually reach you are the ones listed above — not us.

Where this falls short

A protocol that only lists its strengths is marketing. These are the known weaknesses, and we would rather you hear them from us.

  • It matches words, it does not understand. Detection is a hand-written list of patterns, not a trained classifier. It misses distress that is phrased unusually, and it sometimes fires on people who are quoting a song, describing a film, or talking about someone else. Our own design standard calls for a dedicated crisis classifier; we have not built one, and until we do this is a heuristic doing a serious job.
  • It only works in English. Every pattern is an English phrase. If you write to Elafi in another language, the banner will very likely not appear at all — the resources are still reachable from the app and from this page, but nothing will surface them for you.
  • The journal and the couples space are weaker than the chat. Both show the same resources and both raise the urgent tier, but they assess only the entry or message in front of them rather than reading across recent turns — and the check they run is slightly narrower even on a single message, so a few phrasings that reach the urgent tier in the main conversation will not reach it here. Neither keeps the banner the way the main conversation does either — in the journal it lasts until you navigate away or reload, and in the couples space one ordinary follow-up message clears it, with no pill left behind. Only the main conversation keeps a referral on screen for the rest of the day. In the couples space there is a further gap: when both partners write before a reply arrives, their turns are sent together, and our server-side check cannot always tell whose words are whose — so a reply may be shaped by something the other person wrote.
  • A very heavy day can still hit a limit. Elafi has a daily fair-use cap on messages, set well above ordinary use. A hard moment is given room past it — the cap does not close on a conversation that looks like a crisis — but that room is finite rather than unlimited, and past it Elafi does stop replying until the following day. The resources on this page never depend on it.

If something here is wrong

If a number has changed, a line has closed, or Elafi handled a hard moment badly, we want to know — that is the only way the list above stays true. This protocol is reviewed whenever the resources are re-verified.