Work in progress. Crocodile is still being built. Some of what this site says about security and features describes where we're heading and may not be true yet.Some claims here may not be true yet. See the latest status

Built so nobody has to be trusted

Crocodile is designed so that servers, networks and even the friend who hosts a room can't read what you say. This page lists what's protected, what isn't, and how.

What's protected

Your voice and video

Only the members of the call can hear it, or see your camera and screen. It's encrypted on your computer, frame by frame, before it leaves.

Your messages

Only the members of the space or DM can read them, now and later. They're stored on members' devices, never on a server.

Your identity

Only your own devices and whoever holds your recovery key have it. It's kept encrypted with your system's keychain.

What servers see, and what they don't

No system that connects people is free of metadata. Here's exactly where Crocodile draws the line.

Servers never see

  • What you write, say or show: content is end-to-end encrypted, also when you use the opt-in relay or mailbox.
  • Your phone number, email or password: Crocodile doesn't ask for any.
  • Your friends list and blocks: they're encrypted so only your devices can read them.
  • Who sent you a friend request: requests travel as sealed notes that don't name their sender.

Servers do see

  • Public profiles, space names, channel lists, memberships and device names.
  • Who's online and who's in which voice room.
  • While you're connected: whose presence you follow, whom you send a friend request or answer to, and whom you start a DM or call with. Your server doesn't keep that or pass it on, except that to follow profiles live it asks the servers that hold them to watch them, so those servers see which user and space ids it follows while you're connected and for up to five minutes after. They aren't told which user asked, but on a server with one user that is that person.
  • IP addresses, which are needed to connect apps to each other. The host of a room also sees its members' IP addresses.

Who could try, and what they'd get

A threat model lists possible attackers and what each one can and can't do. This is a summary of ours.

AttackerCanCannot
Someone watching the networkSee which IP addresses you talk to, and how much.Read content or metadata, or tell what you say from packet sizes.
Record now, decrypt laterStore your encrypted traffic for years.Decrypt it with a future quantum computer: every key exchange is post-quantum hybrid.
A malicious serverLie about who's online, delay or drop connections, keep logs of metadata.Read content, forge profiles, spaces or invites, swap someone's keys, or get between two apps.
A relay or mailbox operatorSee sizes and timing of relayed packets, or who left how many messages for whom.Read, change or forge them.
The friend hosting a roomSee members' IP addresses and who's speaking when. For video: who has a camera or screen on, who watches which stream, the size and timing of video frames, which show roughly how much moves (video has no constant bitrate), and the picture size.Read more than any member can: their relay only forwards encrypted packets. They could still mislabel who is speaking, or hold frames back or deliver them late.
Someone who left a spaceKeep what they already saw.Read anything said after they left: keys are replaced when someone leaves.
Someone who steals a keyWith a message key: read its current 30-minute period. With a device's prekey secrets: read what was sealed to that device and recorded along the way (by a host, relay or mailbox operator, say) in the roughly eight days before (longer if the device was switched off).With a message key alone: read earlier messages (deleted keys) or later ones (fresh keys).
Someone with your unlocked computerEverything you can. Use disk encryption, and remove a lost device from another one.Keep using the app as shipped once you've removed the device: it stops and wipes itself when it next connects. A device that stays offline or has been modified keeps what it has and can still sign as you.

The cryptography

Well-studied building blocks from audited, open-source libraries (@noble), combined in standard ways.

Ed25519Signatures on every record, message and connection offer
X25519 + ML-KEM-768Hybrid post-quantum key exchange for every pairwise secret
AES-256-GCMEncryption of messages, voice frames and server connections
HKDF-SHA256Key derivation and the forward-only ratchets
DTLS-SRTPWebRTC's encrypted transport underneath
Padmé · CBR OpusHiding message lengths and speech patterns

Known limits

We'd rather you hear these from us.

  • No independent audit yet. The design and code are public, but no outside firm has audited them. Until one has, don't rely on Crocodile where your safety depends on it.
  • Devices share one account key. Removing a device makes the app on it stop and wipe its copy, and stops new messages and keys from reaching it. That is the app's own doing: a device that stays offline or has been modified keeps what it has, and since every device holds the account key, it could still sign as you. Per-device signing keys are planned.
  • Some metadata is visible. Servers see memberships and, while you're connected, whom you contact. Less of it is on the roadmap.
  • Group keys. Crocodile uses ratcheting sender keys rather than MLS. A key is replaced every 30 minutes, but it travels sealed to a device prekey that is kept about eight days (longer if the device was switched off), so forward secrecy against a stolen prekey lasts that long. Replacing keys also costs more as a group grows. MLS is planned for very large spaces.
  • Key checks are manual. You can compare safety numbers with a friend, but the app doesn't prompt you to.

Found a problem?

Please report security issues privately through GitHub, so they can be fixed before they're public.