I built a self-hosted dashboard for keeping tabs on AI agents (Claude Code, OpenCode, Codex, etc.) while they work. It's a Next.js app backed by SQLite that runs on a headless Mac Mini and is accessed over Tailscale. Agents POST status updates to a REST API. You get a drag-and-drop Kanban board with 6 status columns, parent/child task hierarchy (3 levels deep), comments, sprint management, velocity tracking with burndown charts, and 7 analytics chart types. Messages are encrypted at rest with NaCl secretbox. The whole thing binds to 127.0.0.1 — nothing touches the open internet. Tailscale Serve handles HTTPS. No Postgres, no Redis, no Docker. Just SQLite and a Node process. It's deliberately passive: the dashboard shows you what agents are doing, it doesn't orchestrate them. No auto-retries, no task reassignment, no unsupervised autopilot. You still have to check in. What I want to build next: a unified message queue so any IDE terminal can push events without custom hooks, SSE to replace polling, and a proper notification pipeline for blocked tasks. |