confidant.Architecture Overview

Use powerful AI without handing over your data

When you message an AI provider today, that company can read every word. Confidant closes that gap. Your messages are encrypted before they leave your device, and they are only ever unlocked inside a sealed hardware box (a Trusted Execution Environment, or TEE) that your device checks and verifies first.


The problem with plaintext AI

AI = No privacy

Nothing is private

Every medical question, financial decision, or private thought you send to ChatGPT, Claude, Gemini, and others sits in plaintext (unencrypted, readable text) on that company's servers. It is not just visible. It can be pieced together into a profile of who you are and what matters to you.

More adoption

And it's only growing

People and companies keep adopting AI out of fear of falling behind. To make it work well, we hand over more context about ourselves and quietly lower our privacy standards. That trade should not exist. It erodes your ownership of your own data and keeps you from using AI for the moments that matter most.

No defense

ZDRs aren't enough

Many companies sign zero-data-retention (ZDR) agreements, trusting that this protects them. A ZDR is only a promise on paper. If a provider can technically read your data in plaintext, one engineer opening the wrong file, deal memo, patient record, or line of source code, can do damage no contract can undo.


How it works

Here is the path your message takes. Your device connects straight to the server, with no middleman, and checks its hardware receipt (attestation, a signed proof of what the server is running) before sending anything. Inside a sealed hardware boundary, the server can search the web and run the AI model, but only there, in memory nobody, not even us, can reach from outside. The answer is encrypted again before it ever leaves that boundary.

Your ClientChecks the hardware receipt first
direct connection
Sealed hardware boundary (TEE): Intel TDX + NVIDIA Confidential Computing
AttestationSigned proof the hardware is genuine
Web SearchRouted through privacy-first engines
LLM InferenceRuns the model; never leaves as plaintext
encrypted
ResponseOnly your device can decrypt it

What Confidant gives you

Compliance-Ready

Attestation gives you proof, not a promise, of how your data is handled. Your auditors can check that proof themselves, which makes it easier to meet data protection requirements.

Zero Data Exposure

Your prompts and answers never leave the encrypted channel. No intermediary, not even us, can read them. We do not store your name, your passwords, or your IP address. You sign in with a passkey, nothing else.

Provider-Agnostic Search

Your search queries are only decrypted inside the attested enclave, so Confidant, as the operator, cannot read them either. We route through privacy-first providers like Brave and DuckDuckGo, and we can swap them without changing what you trust.

Production-Grade

Confidant can route between multiple AI models, and every server proves its own hardware identity. Builds are reproducible so anyone can audit them. Every disk image is checked on every read. It scales from a single machine to a fleet.


Why we can't read your messages

Not won't be read. Can't be. Three properties of the encryption make that true.

Forward Secrecy

Past data stays protected

Your encryption keys change constantly, and old keys are destroyed for good. Even if a key is compromised years from now, whatever you sent today stays unreadable. Your past conversations are sealed permanently.

Self-Healing Sessions

Breaches can't persist

Fresh encryption material is mixed into every session as it runs. If someone manages to steal a key, the very next exchange locks them back out automatically. A compromise is contained to a single moment, never the whole conversation.

Hardware-Bound Keys

Interception is impossible

The server's encryption key is baked into its hardware attestation report. Your device checks that the key it is using matches the key in that report, proving you are talking to the exact machine that was verified. Man-in-the-middle attacks are ruled out by math, not policy.


How your history stays private

Confidant can sync your conversations across your devices without ever showing them to our servers. Everything is encrypted on your device before it leaves, and only devices that share your passkey can unlock it.

End-to-End Encrypted

We never see your conversations

Every message and title is encrypted on your device with its own key (AES-256-GCM) before it is sent anywhere. Our server only ever stores scrambled ciphertext. Even with full access to that database, your conversations stay unreadable without your passkey.

Passkey-Derived Keys

Your passkey is the master key

Your passkey produces a private output that your device turns into a master key (using HKDF-SHA256, a standard key-derivation method). Each conversation gets its own key, locked under that master key. Only a device holding your passkey can unlock them.

Same-Ecosystem Sync

Sync via iCloud or Google

Passkeys already sync between your own devices: iCloud Keychain on Apple, Google Password Manager on Chrome and Android. Any device with your synced passkey can decrypt your conversations.

How It Works

Passkey
PRF Output
Derive KEK
(HKDF-SHA256)
Per-Thread DEK
(AES-256-GCM)
Encrypt
Locally
Push Encrypted
to Server
Pull & Decrypt
on Other Device

How you can verify it

You do not have to take our word for it. Each link below is something you or an independent auditor can check yourself, from the code we publish to the message on your screen.

Reproducible
Build
Immutable
Disk Image
Hardware
Measurement
Intel-Signed
Attestation
Client
Verification
Encrypted
Channel

How it compares

CapabilityStandard LLM APISelf-HostedConfidant
Data encrypted in transitTLSTLSNoise protocol
Data invisible to providerProvider sees plaintextAdmin sees plaintextOnly in TEE memory
Provable server integrityHardware attestation
Resilient to key compromiseContinuous key rotation
Provider-agnostic searchLocked to provider's searchTypically one providerPrivacy-first engines
No PII requiredPasskey-only auth
Zero operational overheadFully managedFully managed

What's next

TodayShipped

Verifiable private inference

  • Intel TDX CPU enclave with NVIDIA Confidential Computing for GPU inference
  • Noise-protocol end-to-end encryption with forward and post-compromise secrecy
  • Reproducible builds, dm-verity disk images, multi-model routing
  • Per-node hardware attestation, fleet scaling, multi-region deployments
  • Desktop and iOS clients with passkey-only authentication
NextBuilding

Enterprise readiness

  • Integrations with internal enterprise tools
  • Data sovereignty rules to prevent information leakage
  • Pursue SOC 2 compliance
  • Expanded catalog of supported frontier models
  • API key access for programmatic integration
BeyondResearch

Cryptographic private search

  • Cryptographically private search via PIR
  • Providers learn nothing about your queries, by construction
  • Data privacy and sovereignty without hardware trust