Tellwhen

AI alerts for anything on the web

creator · built solo · 2026

Why it exists

Most AI waits for you to ask. The things worth knowing don’t: news that moves your stocks, a new paper in your field, an opening at the company you want, a price that finally drops. Checking for them every day is the chore.

What I built

Tell it what to watch in plain words, and it’ll email you the moment it happens. One email sets it up: no app, no account. I built all of it, the service that watches and the site.

Try Tellwhen (opens in new tab)

Today

Live at tellwhen.app (opens in new tab), and free for the first 100 people.
It emails often within 15 minutes of it happening.
Anything you can put in a sentence: news, a page, or a number on a page.
Every email says when the news first came out, and links its sources.

aAskone email
bWatch3 loops
cDecideAI proposes
dSendcode decides
eStoreone database

Follow an alert through the system

One always-on worker runs its loops side by side: one reads the inbox, three watch, one sends, and three keep house. An AI proposes what to set up and what is worth an email; plain code applies it, keeps the limits and sends each email once. Every loop works from one database, which is also the queue.

Point at any part, or play the walkthrough, to see what it does and the rule it keeps.

The alert this walkthrough follows, emailed Oct 21, 10:31 pm ET

DeepSeek raises $15B in one of China’s largest-ever funding rounds
DeepSeek, the Hangzhou-based AI startup, has raised about 100 billion
yuan ($15 billion) in its latest funding round, one of the largest …

Judged, read at its source, sent once. On tellwhen.app’s share card.

14 parts

The system, following an alert. One email sets it up (a). Three loops watch the news and the pages asked about (b). An AI judges each story against the person’s words and a writer reads the source; a changed page has a check of its own (c). Plain code keeps the limits and sends each email once (d). Every loop works from one database, which is also the queue (e).

Every part, as text

Your email (plain words). You email Tellwhen what you’re waiting for, in your own words. No app, no account.
1. a template to start: What to watch, when it’s worth an email, how fast, and what never to send.
2. change it by reply: Add, drop, pause or stop an alert in plain words.
3. read in seconds: The inbox is checked every five seconds; a failed email is retried, then answered.
The rule it keeps: Every inbound message is stored once, so a retry never acts on it twice.
Sender checks (before any AI). Plain code screens every email before an AI reads a word of it.
1. forged senders: The sender’s domain must pass its own check, or the email is dropped unanswered.
2. machines: Auto-replies, bounces, mailing lists and no-reply senders are skipped.
3. a daily limit: Each sender gets a set number of emails acted on a day.
The rule it keeps: A forged or automated email is dropped before any AI sees it.
Understand (AI proposes). An AI reads the email with the person’s alerts and the thread, and returns a structured decision.
1. a fixed shape: Add, change, pause or remove. An email not about alerts changes nothing.
2. quoted words: Removing, pausing or replacing needs the person’s own words, checked by code.
3. a checked reply: The reply is written from what code did, and falls back to fixed words.
The rule it keeps: Up to five alerts per person, and none that would follow a private individual. Code enforces both, not the AI.
Watch plan (1 to 3 checks). Each alert gets a plan: what counts, what to skip, how fast to check, and where to look.
1. three tools: Search the web, re-read a page, or read one value: usually one to three checks per alert.
2. your speed: Your words set it, from “the moment it happens” to a daily roundup.
3. a hard rule: For a number, code checks it: above, below, rises, drops, changes, or a status.
The rule it keeps: Every page in the plan is read once before the reply promises to watch it.
Page watch (pages, values). The pages and numbers a person asked about, read again at the alert’s pace.
1. its own pace: Every 15 minutes to weekly, as the plan set.
2. only what is new: Only new or changed lines go on; a half-loaded page is skipped, not compared.
3. a new baseline: A page that suddenly doubles, or is mostly new, becomes the baseline, with no email.
The rule it keeps: After five failed reads in a row, the watch pauses and tells the person.
Fast lane (100+ feeds). The feeds where news breaks first: newsrooms, regulators, wires and research feeds.
1. every minute or two: Most feeds are read every one to two minutes.
2. an alert’s own feeds: Feeds found for one alert, such as a company’s own, are read every 90 seconds.
3. an alert’s own searches: Its news searches run every 5 to 30 minutes, by urgency, and sweep the last day every hour.
The rule it keeps: Each new story is matched against every alert at once, inside the database, so the cost doesn’t grow with each person.
Wide net (web search). A web search for each alert, for what the feeds don’t carry, including the plan’s search checks.
1. at the alert’s pace: Every hour, three hours or twelve, by how urgent the alert is.
2. a baseline first: The first search only records what is already out. Nothing is sent from it.
3. an age limit: Results older than the alert’s age limit are marked old news by code.
The rule it keeps: The judge sees old news only to date a story, and can never pick it.
Change check (its own email). A changed page has its own step: an AI decides whether the change matters.
1. the new lines: It sees what changed, against the last full read of the page.
2. a number, proved: Quoted from the page, checked by code, then read again by a stronger model.
3. its own email: It writes the email itself, after checking each picked item’s own date; no judge.
The rule it keeps: Whether a number crossed its line is decided by code, after the second read, never by the AI.
First check (drops off-topic). A quick first check drops what is clearly off topic from the fast lane’s matches.
1. no randomness: Set to answer the same way every time for the same story.
2. fast lane only: The wide net’s results skip it and go straight to the judge.
3. duplicates: In both lanes, a story too close to one already sent is caught by a similarity score.
The rule it keeps: It only drops stories; whether one is worth an email is the judge’s call.
Judge (worth an email?). An AI weighs every story against the person’s own words: who, what happened, and how big.
1. reason, then verdict: It says why, then decides: email, not important, off topic, duplicate or old news.
2. no list of what counts: Nothing about what counts is hard-coded. The bar is the person’s own.
3. it remembers: A turned-down story stays down unless something new removes the reason.
The rule it keeps: A story carried only by social posts is never sent: code enforces it after the judge.
Writer (reads the page). An AI reads the source page itself, confirms the story and writes the email, or declines.
1. dates the event: The day it happened, not the page’s date, so old news on a fresh page isn’t sent.
2. scheduled news: Results and other scheduled news wait for the company or an established outlet.
3. a short email: The news in a sentence or two, with its sources and when it was first reported.
The rule it keeps: A story with no page it can read, that no established outlet carries, is held back.
Outbox (sent once). Every email, alert or reply, waits here with a unique key and leaves at the mail service’s pace.
1. limits: At most 40 alert emails per person a day; an urgent story may add two per alert.
2. replies first: Replies go first, with room kept for them; no daily cap or budget holds one back.
3. sent once: The key goes with the email, so a retry after a crash never sends it twice.
The rule it keeps: More than 500 automated tests guard the code, and new instructions for the AI are tried on real past stories first.
Delivered (with a receipt). The email arrives in the person’s inbox with its sources and when the news was first reported.
1. its sources: Each email links the pages its news came from.
2. first reported: Each news email says when it first came out, such as “First reported today, 4:15 pm ET”.
3. reply to change it: “stop” or “less like this” applies to that alert.
The rule it keeps: The first-reported line names no site, and gives a time only when a second site confirms it.
Database (and the queue). One database holds the alerts and plans, every story seen, every verdict and the outbox. It is also the queue.
1. claimed, not shared: Each loop claims what is due, so two copies of the worker never take the same item.
2. safe to repeat: Unique keys and per-person locks make every step safe to redo after a crash.
3. budgets: Daily budgets for search, AI and email pause work at the limit, but never replies.
The rule it keeps: A status check fails when any loop falls behind, and an alarm watches it.

How it’s built

AI proposes, code decides

The model judges what’s worth telling you. Code enforces the limits, blocks stories you were already sent, and sends each email once.

Your words set the speed

Say “the moment it happens” or “a daily roundup is fine”. News feeds are read every one to two minutes, each alert’s own news searches run every 5 to 30 minutes, and a wider web search and watched pages fill the gaps.

No keyword lists

Every story is judged against your own words.

Proof before an email

It reads the source page and dates the event itself, not the page. A watched number must be quoted from the page and checked again.

Tested before it ships

More than 500 automated tests. Changes to the AI’s instructions are replayed on real past stories before they ship.

What it can’t see

It sees only public news feeds, web search and public pages. Pages behind a login are out of reach.