Trust center · Security

We built a server that can't betray you.

Trust is cryptography, not policy. Here is what SecureComs protects against, what it does not, and how you verify it.

What we protect against

  • • Breached relay or database: the attacker gets ciphertext, no plaintext.
  • • Rogue admin or bad dependency: same result, opaque bytes and no keys.
  • • Membership churn: MLS re-keys on join/leave; old keys do not decrypt new messages.
  • • Tampered history: edit a past record in place and the hash chain fails verification.

What we don't cover

  • • Compromised endpoint: if your device is owned, pairing + safety-number checks are your defense.
  • • Lost recovery key: you hold it; we never can. If lost and all devices are gone, recovery is not possible.
  • • FileSafe after total device loss: the recovery key does not restore vault access. A vault owner who loses every device that holds the vault loses access to its files, so link a second device.
  • • Metadata we must route: channel membership and delivery timing are visible to route; content is not.
  • • Calls with an agent on them. Agent voice runs through the SecureComs voice relay on our servers, which hears and transcribes the call, so those calls are not end-to-end encrypted. To hear and speak, the relay sends call audio to our speech providers, Deepgram (speech to text) and Cartesia (text to speech). The app says so on the call screen.
  • • The Help assistant is outside end-to-end encryption. It has no access to your encrypted messages, files or keys; it reads account details such as your plan, usage and device list, and what you type to it is sent to Anthropic.
  • • FileSafe structure. Our server sees the shape of your folder tree, file sizes, timestamps, who uploaded each file and who has access. Never names or contents.
  • • A ledger rebuilt by a database administrator: anyone with database write access, ours included, can recompute the whole chain so it still verifies. Export the ledger and keep your own copy.
  • • A room member running a modified app. The server refuses removal of the enclave owner and of peer admins, but it never sees inside encrypted membership changes, so a member with a modified client can still disrupt a room they belong to. The owner can re-establish the room. (Your organization's owner-admin can still revoke an account outright.)

Crypto choices, plainly

Group E2E

MLS RFC 9420 with TreeKEM, forward secrecy, post-compromise recovery

Key exchange

X25519 · Signatures ed25519

Backup

Zero-knowledge. The recovery key is held by you, sealed client-side under HKDF(seed). Covers your account and message history, not FileSafe.

Pairing

13-char safety number (Crockford blake2b of ephemeral X25519 pubkey)

The blind relay

The relay stores ciphertext and routes it. No plaintext, no search, no decrypt. That is the contract. The Agent Connector on your side does key generation, rotation, and room re-keying automatically. Humans verify devices with a short safety-number check.

// What the server stores

{ ciphertext: "bA9fX2N…9kLm4pQ==", kid: "mls-epoch-42" }

Tamper-evident ledger

Every message and admin action lands in a hash-chained audit ledger you can export. Edit, delete, or reorder a past record without rebuilding every record after it, and verification fails. Two things it does not catch on its own: records removed from the newest end, and a chain rebuilt by someone with database write access, ours included. Tamper-evident, not tamper-proof, so export and keep your own copy. Retention scales with plan: Free 30 days and Business 365 days.

ExportableHash-chained

Calls, files and agents, plainly

Messages are only part of it. Here is exactly what is end-to-end encrypted when you call, share a file or bring an agent into the conversation, and what is not.

Calls between people

1:1 calls and people-only group calls are end-to-end encrypted. Each device derives the media key from the conversation's MLS group; the key is never sent over the network, so the media server forwards frames it cannot decode. If a browser can't encrypt media this way, the call does not start. It never quietly falls back to unencrypted audio. In a room where the SecureComs voice relay is still a member, calls between people don't start, and the app explains why.

Calls with an agent

Talking to an agent by voice works differently. When an agent joins a call, the SecureComs voice relay joins that conversation and stays in it, so the conversation's calls and messages are no longer end-to-end encrypted. SecureComs runs the relay on its servers, and it sends call audio to our speech providers, Deepgram (speech to text) and Cartesia (text to speech). An admin has to read that and confirm it before agent voice is turned on, and the call screen tells everyone on it that the call is not end-to-end encrypted.

Call transcripts

On a room call with an agent, the voice relay transcribes what it hears and posts the transcript into the channel. Everyone in the channel when the call ends can read it. Because the relay heard the call on our servers, a transcript like this is a disclosure, never an encryption claim, and the app labels it that way.

FileSafe

File names and contents are encrypted on your device before upload, and each enclave's vault is its own encrypted group. What our server does see: the shape of the folder tree, file sizes, timestamps, who uploaded each file and who has access. That is the trade-off for folders that sync across your team.

Backup & device trust

Zero-knowledge backup

The backup blob is sealed client-side; the server holds ciphertext and a recovery envelope sealed under your recovery key. Only the Clerk-authenticated identity owner can touch them. Agent API keys and device JWTs get 403. It restores your account identity and message history. It does not hold FileSafe vault keys: a new device reaches the vault only when a vault owner's device that already holds it admits the new one.

Device pairing

Link a new device with a 13-character code derived from an ephemeral X25519 pubkey. A substituting server produces a mismatching code, so the human aborts. That is the MITM defense.

SecureComs supports the browser matrix documented on this site.