How I Built a Multi-Model AI Council That Runs on a Mac Mini

I run 4 AI agents (Claude Code, OpenClaw, Hermes/DeepSeek, LM Studio) on a single Mac Mini M4 with 32GB RAM. They share memory through Obsidian + ChromaDB, communicate via ACP bridge, and delegate tasks using a tiered hierarchy. Here's what actually works and what breaks.

The Stack

• Orchestrator: DeepSeek V4 Pro (API) — plans, delegates, communicates
• Architect: Claude Code — design decisions, code review
• Workers: LM Studio with qwen3.5-27b — heavy coding, research
• Memory: Obsidian vault + MemPalace (ChromaDB) — shared across all agents
• Communication: acpx bridge — npx acpx claude exec "message"

The Hard Part: RAM At 32GB

32GB minus ~7GB for macOS = 25GB usable. One large model (~22GB) leaves almost nothing for subagents. The solution: cloud orchestrator (DeepSeek) + local workers. The orchestrator costs ~$3/month in API tokens. Workers cost electricity (which runs anyway).

Arena Council: Multi-Model Voting

When one model isn't enough, I run 3-5 local models in parallel and let them vote. Each model gets the same query, responses are scored, and the best one wins. It's like having a panel of experts instead of one.

# 5 models, 1 answer
council = ArenaCouncil(models=["qwen-27b", "qwen-35b", "mistral-7b", "llama-8b", "qwen-9b"])
result = council.ask("Should I refactor this monorepo?")
# → 3/5 say yes → majority wins

All local, all free (beyond electricity).

What Broke (And How I Fixed It)

1. Sandbox mode — killed symlinks to external drive. Disabled it.
2. LM Studio kernel panics — loading 2 large models simultaneously on M4 GPU. Now max 1 large model.
3. Doctor --repair — overwrote config files. Now backup before any repair.
4. Lobotomy — agent overwrote its own personality file. Added file-level write protection.

Full story + code: github.com/nerua1

---

If this saved you time: ☕ PayPal.me/nerudek