Tern
Private beta · iPhone

Your people.
Your device.
Your data.

Tern is an encrypted messenger whose intelligence runs on your phone — not in someone's data centre. Real conversations, real encryption, and an assistant that still works on a plane.

Shipping, not slideware

What works today

Everything below runs on real phones in a private household beta. Where something isn't finished, this page says so — that is the whole point of building a trust product in the open.

🔐

Encrypted roomslive

Create a room, share its code, talk. Newer rooms run on MLS — the IETF messaging-security standard — so every member has their own keys, past messages stay safe if a device is later compromised, and the group re-keys whenever membership changes.

Catch me uplive

One tap summarises what you missed: decisions, questions nobody answered, things asked of you. It reads your messages on your phone. Today's version is rule-based; the on-device AI model comes next, and the app always states which produced a summary.

📷

Photos & voice noteslive

Snap a photo or hold to record. Photos are resized and re-encoded on your phone, which strips the GPS data most apps forward without asking. Media is sealed and streamed in pieces like everything else.

🔔

Notifications that reveal nothinglive

Our server cannot read your messages, so a notification can only say "you have new messages" — your phone fetches and decrypts the content itself. There is no preview text on the server, because there can't be.

👥

Invites without harvestinglive

Invite someone from your contacts while your address book stays on your phone: iOS shows the picker and hands us only the person you tapped. It is also why there's no "who else is here" list — we would have to upload your contacts to build one.

💬

The everyday thingslive

Reactions, editing, typing indicators, read receipts, unread counts, delete for everyone and delete just for you. Even read receipts are encrypted state between your phones — our server never learns who read what.

Under the hood

The cryptography, stated plainly

We use established, standardised primitives through audited libraries and write nothing of our own. Anyone evaluating a privacy claim deserves the specifics rather than a padlock icon.

Group encryption
MLS (RFC 9420) via OpenMLS, ciphersuite X25519 · AES-128-GCM · SHA-256 · Ed25519. Every member holds their own keys; the group re-keys on each membership change, which gives forward secrecy and post-compromise security. live
Transport
Sealed payloads travel inside an authenticated ChaCha20-Poly1305 envelope over TLS 1.2+. The relay sees ciphertext on an unlinkable topic — never content, authorship, or document structure. live
Key storage
Message keys and group state live only on your devices and survive app restarts locally. There is no server-side key escrow, so there is nothing for us to hand over. live
Media & metadata
Photos and voice notes are sealed and chunked like messages; photos are re-encoded on-device, dropping EXIF/GPS. Read receipts and typing hints are encrypted state between your phones. live
Post-quantum
Hybrid X25519 + ML-KEM (FIPS 203) key agreement, so traffic recorded today cannot be decrypted by a future quantum computer. The decision is written and approved; the swap lands when our MLS library ships production support, and the codebase is already structured so it touches one layer. planned
Independent review
An external security audit — including post-quantum readiness — is a hard gate before Tern opens beyond invited testers. Our cryptographic decisions are recorded as numbered architecture records so an auditor can read the reasoning, not just the code. planned
Why switch

What changes when you're the customer, not the product

This compares Tern's design commitments with practices that are common — and publicly documented — across ad-funded platforms. It describes an industry norm, not any single company.

Typical ad-funded appTern
Business modelAdvertising — your attention is what's soldSubscriptions, planned — you pay, so nobody else has to be served
Where the AI runsCompany data centres, processing your contentOn your own device; your content doesn't leave it
What the server can readMessage content, for features and scanningNothing — it relays sealed bytes on anonymous topics
Your contactsAddress book uploaded to find other usersNever uploaded; you send invites yourself, from your phone
Read receiptsMetadata the server can seeEncrypted state between your devices
Ad profile of youBuilt from your activity and interestsNone exists — there are no ads to target
Session designInfinite scroll, streaks, autoplay are commonNatural stopping points; no streak mechanics, ever
ModerationOften opaque; appeals vary widelyDesigned so every action carries a written reason and a right of appeal
AI labellingInconsistentEvery AI surface states how it produced what you're reading

The "typical" column summarises practices documented in regulatory findings and companies' own disclosures; individual apps differ, and some do better on some rows. The Tern column mixes shipped behaviour with commitments for features still being built — the roadmap below keeps them separate.

Trust layer

Built for the rules that are arriving, not the ones that expired

Regulators across the EU, the UK and a growing list of US states are converging on the same demands: explain your moderation, label your AI, minimise what you collect, and protect minors. We treat those as design inputs — decided before a feature is built, not patched on afterwards.

Moderation you can contest planned

Every enforcement action is designed to carry a written, human-readable reason and a route to appeal, in the shape the EU's Digital Services Act expects. The pipeline — triage, statement of reasons, appeal — is already built and tested in our core; the community surfaces that use it come with Phase 3.

Age assurance, without an ID database planned

New York's SAFE for Kids Act, the UK's Online Safety Act and several US state laws require platforms to establish age before exposing minors to certain features. Our design keeps that private: a device-held or third-party attestation of "over 18" or "under 16" rather than uploading your passport to us. We would rather hold no identity documents at all than promise to guard them well.

Labelled AI live

Every AI surface states how it produced what you're reading — today that means Catch Me Up telling you it used rules rather than a model. That transparency is what the EU AI Act asks for, and we shipped it before anyone required it of us.

Data minimisation by default live

Ranking signals and interaction history stay on your device. We do not upload contacts, we do not build ad profiles, and the server stores no message content — so there is very little for a breach, a subpoena, or a change of ownership to expose.

A feed you can opt out of planned

When community feeds arrive, a chronological, follows-only option is one tap away and always available — the non-profiling choice EU rules now push the industry toward. No streaks, no autoplay for minors, and a scroll that ends.

Transparency reporting planned

Moderation actions, recommender parameters and systemic-risk metrics are being designed to be queryable for regulator and researcher reporting from the start, rather than reconstructed under deadline.

Where we are — honestly

Built in the open, shipped when ready

Tern is in private beta: a working encrypted messenger on real iPhones, not a mockup. It is also early — one platform, invitation only, and the community half is still ahead of us. Our architecture decisions and benchmarks are written down as we go, because "trust us" isn't a privacy policy.

56 tok/sternary model on a desktop CPU — reading speed, no GPU
20–25 tok/smeasured on iPhone 13 and iPhone 16, single thread
0 bytesof your message content our servers can read
Phase 1 · done

The hard bets, proven

  • Ternary AI running at reading speed on ordinary phones
  • Local-first encrypted data that survives offline and merges cleanly
  • A relay that carries sealed bytes and nothing else
Phase 2 · now

Private beta on iPhone

  • Encrypted rooms with per-member MLS keys
  • Photos, voice notes, reactions, receipts, notifications
  • On-device "catch me up" summaries
  • Next: the AI model inside the app, voice transcription, Android
Phase 3 · next

Communities & the trust layer

  • Channels, roles and admin tools inside a space
  • An AI co-moderator that learns a community's own norms
  • Moderation with written reasons and appeals, wired to the UI
  • Privacy-preserving age assurance; transparency reporting
  • Discovery, and a feed you can always set to chronological
Later · gated

Hardened & extended

  • Post-quantum hybrid key agreement once the library lands
  • External security audit before any public release
  • Wider regional coverage as age and safety rules take effect
  • Agents with their own cryptographic identity — scoped, revocable
  • Payments only after identity, audit trail and licensing work are done — deliberately deferred, in writing, with no model ever in the money path
The business model

How we plan to make money

Subscriptions and paid community tools. That single choice does most of the privacy work: when you pay for the product, there is no advertiser to profile you for, no reason to engineer a feed for time-spent, and no incentive to read anything you write.

Nothing is for sale yet — the beta is free and invitation-only, and payments inside Tern are deliberately deferred until the identity and audit layers can carry them.

Things we've decided not to build

Advertising. An engagement-ranked feed as the only option. Streaks. Uploading your address book. Reading your messages to train models. Payments before the groundwork is ready. Each of those is a written decision with reasons, not an oversight we might quietly reverse.

Private beta

Come test it with your people

Beta groups are small on purpose — households and tight-knit communities who move together, so the app gets tested the way it is meant to be used. iPhone today; Android next.