Conway
A personal AI operating system.
Conway began as a chat-UI experiment and grew over a year of daily use into a personal AI operating system. Ask it anything in plain language and specialized agents handle the work across the systems it connects. It also watches email, transactions, and the calendar on its own, and surfaces patterns worth your attention.
- 8 domain agents
- 5 ambient agents
- 10+ integrations
- ~$50 a month to run, mostly LLM spend
How it works
- Agents for each domain Eight agents cover calendar, reminders, tasks, memory, knowledge, finances, reading, and weather. Agno orchestrates them, with LLM providers switchable at runtime: Claude Haiku routes intent and dispatches tools, and Claude Sonnet handles extraction and summarization.
- Ambient intelligence Domain publishers emit events to a Redis Streams bus. A pattern engine correlates them, and five ambient agents surface insights about finance, calendar, email, wellness, and memory without being asked.
- Real daily load Every Gmail message flows through the event bus and is summarized hourly. Messages from whitelisted senders or flagged as urgent surface in real time, and daily briefings run on a schedule.
- A Trust Ladder for autonomy Agents earn autonomy through demonstrated reliability instead of getting it up front. The model is in place, and most agents still observe and surface rather than act on their own.
- Hybrid retrieval pgvector embeddings with keyword, semantic, and reciprocal rank fusion search across documents, notes, and emails.
- Self-hosted in production A home-lab Kubernetes cluster with Gitea CI, GitHub Actions, and PostgreSQL backups to Cloudflare R2 twice a day with 14-day retention.