Proactive AI Assistant

A self-hosted agent that decides when to interrupt me

I did not want another assistant I had to ask. I wanted one that knew when to speak. It runs on a private server, reads my mail, my calendar and a summary of my spending, and wakes on a schedule I cannot predict. A cheap model decides whether anything matters. A better one writes the message, or throws it away. Most of the time it says nothing, which is the part I am proudest of.

• Least-privilege agent design: models read files, never APIs, and anything scheduled or security-relevant is owned by deterministic code
• Delivery runs through a root-owned systemd bridge with the recipient hardcoded outside the model's reach, so a confused or compromised agent cannot message anyone else
• Tailscale-only VPS with no public ingress, and EU Vertex AI endpoints for data residency
• Behaviour lives in git-versioned Markdown contracts rather than code, so tuning the assistant means deploying words
• Append-only ledger for dedupe and audit
• Consumes the digest API from my finance platform, so the two systems work together

Python, Google Vertex AI, systemd, Tailscale, Telegram Bot API, SQLite