What's new in Cognition
A timeline of features, fixes, and improvements.
Pay for AI access with cryptocurrency. No credit cards, no middlemen.
- Deposit ETH, USDC, or USDT on Ethereum or Base — each user gets a unique deposit address
- Automatic detection — deposits are detected within 30 seconds and converted to USD at live market rates
- OpenRouter API keys — convert your balance into a personal API key with 200+ models (Claude, GPT-4, Llama, Mistral, Gemini, and more)
- Your key, your access — the key works in Cognition or any OpenAI-compatible client
- Credits page — balance card, chain selector, deposit address, transaction history
How it works
- Go to Credits in the sidebar
- Select Base (low fees) or Ethereum
- Send crypto to your deposit address
- Click "Purchase Key" once your balance shows
- Add the key as an OpenAI-compatible provider in Settings
Three new subsystems that turn Cognition into a personal AI operating system.
- Autonomous agents — schedule AI agents to run on intervals (every hour, daily, etc.) with goals, tool access, and alert notifications
- Knowledge graph — encrypted personal knowledge base with typed nodes (documents, bookmarks, code, notes) and relationship edges
- App layer — install sandboxed apps that can access AI, knowledge, and memory through a permission system
Live preview panel for code generated during chat.
- Auto-opens during streaming — when the AI writes HTML, SVG, Mermaid, or markdown, the canvas panel slides open and renders it in real time
- Sandboxed HTML — interactive HTML previews in a secure iframe
- Mermaid diagrams — flowcharts, sequence diagrams, and more rendered automatically
- Multi-artifact tabs — multiple code blocks in one response get tabbed navigation
- Canvas button on historical code blocks — revisit any artifact from past messages
Run models entirely in your browser. Zero network, zero API keys, zero data leaving your machine.
- WebLLM/WebGPU — SmolLM2, Llama 3.2, Qwen 2.5, Phi 3.5 Mini run directly in-browser
- Blue "Device" badge — distinguishes browser-native models from local and cloud
- One-time download — models cached in browser storage, instant on subsequent loads
The AI remembers things about you across conversations — encrypted at rest like everything else.
- Auto-extraction — preferences, facts, and context extracted from conversations
- Categories — preference, context, style, fact, general
- Memory page — view, edit, and delete memories from the sidebar
- Context injection — memories automatically included in chat context
Signal-style proof page — verify the encryption yourself.
- Vault fingerprint — unique identifier derived from your encryption key
- Raw ciphertext samples — see what the server actually stores, with live decrypt buttons
- Database column audit — complete map of encrypted vs. plaintext fields
- Interactive test — encrypt and decrypt arbitrary text in your browser
Connect a Telegram bot for 24/7 AI access and workspace control from your phone.
- Per-user Telegraf bot — bring your own @BotFather token
- Workspace commands — /bash, /files, /ws from Telegram
- AI chat — full agent loop with tool calling, same as the web app
- Server-side encryption — bot tokens encrypted with separate key (JWT_SECRET-derived)
Isolated development environments with terminal, file browser, and AI coding agent.
- Container-per-user — Ubuntu 24.04 with Node, Python, ripgrep, git
- WebSocket terminal — interactive shell via xterm.js
- File browser + code editor — recursive tree, language detection, Ctrl+S save
- AI coding agent — 7 tools (read, write, edit, bash, glob, grep, list) via docker exec
- 3-panel layout — file tree | editor + terminal | AI chat
Zero-knowledge encrypted AI platform with multi-user auth, streaming chat, and agent workflows.
- AES-256-GCM encryption — all conversations, messages, and API keys encrypted client-side
- PBKDF2 key derivation — 600,000 iterations, non-extractable CryptoKey
- Multi-provider support — OpenAI, Anthropic, Ollama, any OpenAI-compatible
- Agent builder — custom agents with system prompts, models, and icons
- Workflow builder — chain agents into multi-step pipelines
- Auto-lock — configurable timeout, key wiped from memory