Self-hosted AI ops on Mac mini: Claude Max + Discord + MCP at $0/month extra

Running a personal AI operations system for 2 months now. Here's the full architecture.

Hardware & Stack

• Hardware: Mac mini M2 (always on, silent, no cloud bill)
• Runtime: Node.js + bash
• Memory: LanceDB (local vector) + SQLite (queues)
• UI: Discord
• AI engine: claude -p headless CLI (included in Claude Max $20/month)

What runs 24/7

1. Reactive Discord chat — multi-turn threads, session memory, RAG knowledge base
2. 30 scheduled cron tasks — standups, market alerts, code audits, log rotation, vault sync
3. 11 specialized agent teams — each has a role, schedule, and output channel
4. Self-healing infrastructure — 4-layer recovery: preflight → keepalive → watchdog → Claude AI diagnosis

The Context Intelligence Gateway

Main problem with 24/7 Claude agents: context window fills up in 20-30 minutes on heavy workloads.

Built a local MCP server (Nexus CIG) that intercepts every tool output, classifies type (JSON/log/table/markdown), extracts signal, drops noise before it hits Claude's context window.

315 KB raw output → 5.4 KB compressed (~98%)

Sessions now run 3+ hours reliably.

Numbers

• Uptime: 99.7%
• Context savings: ~98% on heavy tool use
• Incidents: 14 total, 9 auto-resolved (64% autonomous recovery)
• Extra cost: $0/month

Open source (MIT)

• https://github.com/Ramsbaby/claude-discord-bridge
• https://github.com/Ramsbaby/openclaw-self-healing — autonomous crash recovery
• https://github.com/Ramsbaby/openclaw-memorybox — memory hygiene CLI

Happy to answer questions about any part of the architecture.