Chat apps already measure how active you are, but only for themselves: engagement dashboards, "last seen", streaks that pressure you to post. Group admin bots count messages to rank the loudest members. Check In flips that same signal around: it watches for the moment a friend goes quiet against their own normal rhythm, and turns it into a private, gentle nudge to the people who love them. Nobody else uses silence as a prompt for care.
The data it runs on
Source: message counts per person per week, nothing else. Two ways in. (1) Chat exports: WhatsApp and Telegram both let a member export a group chat as a .txt or JSON file; the app parses timestamps and sender names on the device and throws the text away before anything leaves the phone. (2) A small bot that group members add (Telegram Bot API, Discord bot, Slack app). It records only "sender X posted at time T".
Limits and gaps: WhatsApp has no open API for personal groups, so WhatsApp is export-only and refreshes when you re-export (weekly reminder). iMessage has no export path on iPhone. Some people are naturally quiet in groups but chat 1:1, so a baseline needs at least 4 weeks before any nudge. Holidays, exams and time zones cause false alarms, so the model checks if the whole group dipped first.
Fallback: if no export or bot is possible, the user ticks a weekly "who have I heard from?" list in 20 seconds, and the same rhythm logic runs on that.
1. The mission
Nobody should drift away from their friends without someone noticing. Check In uses the quiet patterns chat apps already see to help you reach out at the right moment, with the right words, and never makes it feel like surveillance.
2. Why people can't resist it
Being a good friend, made easy. Most people want to check in and simply don't notice. One warm nudge turns guilt into action.
The reply is the reward. "Thank you for asking, honestly it's been a lot" is one of the most meaningful messages a person can receive.
A street that lights up. Every check-in lights a window. Your care history becomes visible, without leaderboards or scores.
3. How it works for real
Ingest: parse the export or bot events on device into weekly counts per member. Text is discarded immediately.
Baseline: a simple robust statistic (median and spread of the last 7 weeks per person). No AI needed, no cost.
Notice: flag when this week is far below a person's own median, unless the whole group dipped (holidays) or the person paused noticing.
Classify gently: a fast cheap model (Gemini Flash or DeepSeek via OpenRouter) picks the right tone from metadata only: how close you are (your 1:1 frequency), how long since you spoke, time of year.
Write: a frontier LLM (Claude or GPT) drafts 3 short openers in your own style, learned from a few of your sent messages that you choose to share. No mention of tracking in the draft.
Care mode: if a reply sounds heavy, the app shows listening tips and general crisis-resource wording for the user's country. It never reads the friend's reply unless the user pastes it in.
4. Cost per use
Near zero. Counting and baselines are arithmetic on the phone. One set of 3 drafted messages costs about $0.002 with Claude Haiku or Gemini Flash, and most weeks have zero or one nudge.
Free: 3 groups, nudges, drafts, history. Plus ($2 a month): unlimited groups, family mode (parents and grown children), birthday and anniversary memory, and a yearly "people you showed up for" book.
5. Growth loop
The bot joining a group posts one consent message to every member. Each member learns the app exists, and the ones who like it install it for their own groups. The check-in itself spreads: people who were checked on are much more likely to check on someone else.
6. Risks and guardrails
Consent: every member is told once and can opt out of being noticed with one reply. Opted-out people are never counted, not just hidden.
Privacy: counts only, never message text, stored on device. No "who they talk to" graph. Nothing is ever shown to the friend.
Not a clinical tool: no risk scores, no diagnoses, no "warning" language. The app says "quieter than usual", never "at risk".
Crisis: clear wording to contact local emergency services or a crisis line, shown when someone shares they are struggling.
Abuse: no tracking of partners or exes: a person must be in a shared group, and the app refuses 1:1 monitoring.
7. MVP build order
WhatsApp .txt export parser that outputs weekly counts per sender, on device.
Sparkline list plus the "quieter than usual" rule with the group-dip check.
Claude prompt that drafts 3 openers in 3 tones, with copy and "Open WhatsApp".
Check-in history and the lit-window street.
Telegram bot with the one-time consent message and STOP opt-out.