AI Privacy, Built In

Every AI platform reads your prompts in plaintext. Cognition encrypts everything in your browser before it touches a server. No download required. Just open the app and start chatting.

AES-256-GCM
600K PBKDF2
Nothing to install
The problem

Every AI platform reads your data

ChatGPT, Claude, Gemini, Open WebUI, LibreChat. They all store your conversations in plaintext databases. Your M&A strategy prompt. Your medical question. Your proprietary code. Sitting on someone else's disk, readable by anyone with server access.

Every other platform Exposed

you plaintext Server readable admin, breach, subpoena = full exposure

Server admins, data breaches, and legal subpoenas can expose everything. Your data exists in plaintext at rest.

Cognition Encrypted

you ciphertext Server blind breach = noise. no key, no data.

Encryption happens in your browser. The server stores ciphertext blobs. A full database dump reveals nothing.

How it works

Five steps to zero knowledge

Every piece of data is encrypted before it touches a network. The server never holds the key. Not by policy. By math.

01

Password

You enter a passphrase. It never leaves your browser.

client-side only
02

PBKDF2 derivation

Your password derives a 256-bit master key. 600,000 iterations with a random salt make brute-force attacks take years.

600,000 iterations / SHA-256
03

Master key

A non-extractable CryptoKey held only in JavaScript memory. Not in localStorage. Not in cookies. Not on disk.

Web Crypto API / non-extractable
04

AES-256-GCM encrypt

Every message, title, and API key is encrypted with a fresh random 12-byte IV. Authenticated encryption detects tampering.

random IV per operation
05

Ciphertext stored

The server receives base64(iv):base64(ct) and stores it. That's all it ever sees.

server is a dumb data store
password your device PBKDF2 600K iterations SHA-256 + salt AES-256 GCM mode random IV base64(iv):base64(ct) → server stores this your browser server
Features

Everything you need. Nothing exposed.

A complete AI platform with encryption baked into every layer. Not bolted on. Built in.

Encrypted chat

Every message encrypted with AES-256-GCM before it leaves your browser. Streaming responses, markdown rendering, syntax highlighting. All decrypted locally in real time.

Zero-knowledge
How do I optimize my... Here are three approaches... Show me the code for... encrypted

Multi-provider hub

Connect OpenAI, Anthropic, Ollama, or any OpenAI-compatible endpoint. Switch between models mid-conversation. API keys are encrypted at rest, decrypted only in memory per-request.

Bring your own keys
OpenAI Claude Ollama Custom

Agent builder

Create specialized AI agents with custom system prompts, model selection, and personalities. Start from templates for research, coding, writing, or analysis. Agent configs are encrypted like everything else.

4 templates included
Agent Researcher Coder Writer Analyst system prompt + model + personality

Workflow pipelines

Chain agents into sequential pipelines. Output of one becomes input to the next. Build research-to-analysis-to-summary workflows with live streaming output at each stage.

Multi-step orchestration
Research agent 1 Analyze agent 2 Write agent 3 output all intermediate results encrypted before storage
Docker workspaces

Your code. Your container.

Every user gets their own isolated Docker container. No shared files. No shared environment. Your workspace is completely private — other users can't see or access anything in it. Full dev environment with Ubuntu 24.04, Node 20, Python 3, git, and an AI coding agent. All in the browser, nothing to download.

workspace — cognition FILES src/ index.ts app.ts utils.ts tests/ package.json app.ts 1 import { Router } from 'express' 2 3 const app = new Router() 4 app.get('/', (req, res) => TERMINAL $ npm run dev Server running on :3000 $ _ AI AGENT Add error handling I'll add try-catch to the route handlers... edit_file

Per-user isolation

Each user gets their own Docker container. No shared files, no shared processes. Your workspace is invisible to other users.

Full dev tools

Interactive terminal, file browser, code editor with line numbers and language detection. Real development, in the browser.

AI coding agent

7 tools: read, write, edit, bash, glob, grep, list files. The agent works inside your container, not on the host.

Telegram bot

Cognition in your pocket

Connect a personal Telegram bot to your Cognition instance. Chat with AI, run bash commands in your workspace, and manage files. All from your phone.

Each user gets their own bot. The bot token and API key are encrypted server-side with AES-256-GCM using a key derived from the JWT secret.

  • /start Link your account
  • /workspaces List your workspaces
  • /ws name Select active workspace
  • /bash cmd Run command in container
  • /files List, read, or write files
  • Plain text routes to the AI agent with tool calling.
Cognition Bot Optimize my Dockerfile Multi-stage build will cut image size by 60%. edit_file Dockerfile /bash npm test 12 passing (340ms) Message...
Architecture

Designed to be blind

The server is architecturally incapable of reading your data. Not by policy. By design. It stores ciphertext. It retrieves ciphertext. That's all it can do.

Browser PBKDF2 key derivation AES-256-GCM encrypt Web Crypto API Svelte 5 UI key lives here only ciphertext Server Store ciphertext Retrieve ciphertext SSE proxy SQLite + SvelteKit no key, no plaintext API proxy AI Backends OpenAI Anthropic Ollama (local) OpenAI-compatible encrypts everything sees only ciphertext your keys, your models
Transient API keys: During AI requests, the browser sends the decrypted API key per-request. The server uses it to proxy to the AI provider, then discards it. The key exists in server memory only for the duration of that HTTP request. Never written to disk or database.
Deploy

Deploy in 60 seconds

Single Docker container. SQLite database. Zero external dependencies. Or use the hosted version right now.

Docker

# Pull and run docker compose up -d # Open http://localhost:3000

Hosted

# No install needed https://app.cognitionai.tech # Create vault → add provider → chat

Bring your own keys. Cognition doesn't bundle AI models. Connect OpenAI, Anthropic, Ollama, or any OpenAI-compatible provider.