Six things Latch does — and the constraints behind each one.
Each feature exists because of a thing we believe about how operators actually work. We've listed the belief next to the feature. If we ever drift from one, hold us to it.
- Triage
A ranking model that actually knows your job.
Latch builds a model of you in your first week — who you reply to, how fast, who you ignore, which phrases mean 'I need a decision' versus 'FYI'. After seven mornings of confirmations, it gets >95% of your inbox right.
Signals used: sender history, thread velocity, your last-30-days reply patterns, calendar proximity, explicit asks, deadline language. Not: subject-line keywords. We don't use them.
- Drafts
First-draft replies in your voice.
For every thread that needs a response, Latch writes a draft grounded in how you've actually replied to similar threads. Your phrasing. Your tone. Your sign-off. You edit a sentence and send.
Drafts are scoped to your account and trained on your sent-mail only. No cross-tenant learning. No fine-tuning on your data without your explicit opt-in.
- Summary
One email at 7am. Not 4,000 push notifications.
Every morning, Latch sends one summary: the four threads that need you today, what each one is asking, and a one-tap link to the draft reply. That's it. No app to open. No badge count. No 'check Slack'.
You choose the delivery time and day-of-week pattern. Skipping weekends is the default. You can also pause Latch for a stretch — vacations, sprints, surgery, whatever.
- Edge
Your inbox stays at the Cloudflare edge.
Latch runs on Cloudflare Workers and Durable Objects. Your messages are processed at the closest edge node and never sit on a long-lived server. The model that knows you is a per-account Durable Object — isolated, encrypted, deletable in one click.
Threat model: assume the platform is breached. Per-account isolation means a breach exposes one inbox, not all of them. Encrypted at rest. Audit log of every read.
- Honesty
When Latch is unsure, it says so.
Confidence scores are visible on every triage decision. Below 70% confidence, Latch surfaces the thread anyway and asks. We'd rather waste 30 seconds of your time than archive something that mattered.
Confidence is calibrated, not vibes. We log the post-hoc outcome of every decision and re-fit weekly. The model drifts in your favor.
- Exits
Leave anytime. Take the model with you.
One click exports the model that learned you — your reply patterns, your priorities, your archive history. The DO is deleted. The mail-server connection is revoked. No 30-day grace period, no recovery dark patterns.
Export format: portable JSON. Re-importable into the open-source self-hosted build (coming v1.2).
See it on your inbox tomorrow.
14 days free. Setup takes three minutes. We'll show our work.