Multi-Agent AI Setup Guide

Battle-tested setup guide for a multi-agent AI system on Apple Silicon (Mac Mini M4). Covers: install order (Claude Code first), RAM reality for local orchestrators, shared memory (Obsidian + MemPalace), inter-agent communication, critical pitfalls (doctor destroys config, lobotomy protection, sandbox kills symlinks), and WhatsApp console loop. Load when setting up a new system or onboarding a new agent.

OpenClaw + Claude Code Setup Guide

Real-world lessons from running 24/7 multi-agent AI on Mac Mini M4

This is not the happy-path documentation. This is what actually happens.

---

1. Install order matters: Claude Code first

Why Claude Code before OpenClaw:

OpenClaw is powerful but fragile — it can lobotomize itself (overwrite its own config/personality). Claude Code acts as an independent safety net: it reads files directly, doesn't share OpenClaw's context window, and can diagnose/repair a broken OpenClaw from the outside.

Install order:
1. Claude Code (claude.ai/download or brew)
2. Configure Claude Code memory + skills
3. Install OpenClaw
4. Connect them via openclaw-bridge

If you install OpenClaw first and it breaks before Claude Code is ready, you have no recovery path.

---

2. RAM reality: why cloud orchestrator makes sense at 32GB

The actual constraint

32GB minus ~6-7GB for the OS = ~24-25GB available. That's enough to run one large model at a time — qwen3.5-35b-a3b or huihui-qwen3.5-27b loads fine.

The problem is architectural: if you load a 35B as orchestrator, there's no RAM left for a subagent. You can use a smaller orchestrator to fit a subagent — but then the orchestrator is too weak to actually orchestrate.

The real issue: quality gap

The difference between a local model and Kimi k2.5 as orchestrator is enormous. Kimi offers reasoning-level quality (comparable to Claude 3.5 Thinking) at the cheapest tokens available right now. A local 35B next to it is a different league for planning and delegation.

What works at 32GB

Orchestrator: Kimi k2.5 (cloud, cheap, real reasoning)
Subagents:    LM Studio local 30B (free, good enough for execution)

At 64GB+ you can experiment with a local orchestrator — but you'll likely come back to Kimi once you see the planning quality difference.

Good cheap API options for orchestrator:

• Kimi k2.5 (moonshot) — best reasoning per token right now, handles Polish
• GLM-4-Flash — very cheap ($0.20/M tokens), good for simple routing
• Claude API — expensive but best quality (use subscription via Claude Code instead)

Rule: never run orchestrator locally on 32GB

If you try to run a 35B orchestrator + 30B subagent on 32GB:

1. macOS swap kicks in
2. Response time: 30s → 5 minutes
3. System becomes unusable for anything else
4. Eventually kernel OOM kills the process

Keep orchestrator in the cloud, workers local.

---

3. The doctor problem: NEVER run openclaw doctor --repair

What happens

openclaw doctor --repair (or --fix) overwrites your custom configuration with factory defaults. This includes:

• Your custom AGENTS.md routing rules
• Your skill configurations
• Your channel settings
• Your personality/identity files

This is irreversible if you don't have a backup.

Safe alternatives

# Check health without repair
openclaw doctor

# If you see errors, fix them MANUALLY by reading the error and editing files
# Do NOT pass --repair, --fix, or --force

# If you must repair one specific thing:
openclaw doctor --fix --only gateway-token  # hypothetical, check your version

Before any doctor command: backup first

openclaw backup create
# or manually:
cp -r ~/.openclaw ~/.openclaw.bak.$(date +%Y%m%d)
cp -r /path/to/workspace /path/to/workspace.bak.$(date +%Y%m%d)

---

4. Lobotomy protection: agents that overwrite themselves

The problem

OpenClaw (and any LLM agent) can be instructed — by malicious input, confused context, or its own reasoning — to overwrite SOUL.md, AGENTS.md, or other identity files. This is called lobotomy: the agent loses its personality, routing rules, and memory.

Signs you've been lobotomized

• Agent doesn't remember its own name or capabilities
• Routing rules are gone (routes everything to expensive models)
• Skills not loading
• Duplicate folders appearing in workspace (sandbox/ copies)
• Agent starting from scratch every session

Protection in AGENTS.md

## BEZWZGLĘDNE ZAKAZY
- NEVER edit SOUL.md, AGENTS.md, MEMORY.md without explicit owner request
- If you see another agent (or yourself) overwriting a file → STOP, restore from backup
**SAFE examples:** LM Studio subagent generation, file reads, web search on trusted domains, `capture-idea` to `ideas/`
**NEVER examples:** editing SOUL.md/AGENTS.md, rm -rf, installing >35B models, opening ports, sandbox ON
### Recovery procedure

```bash
# 1. Check git history if workspace is a git repo
git log --oneline -20

# 2. Restore from backup
cp ~/.openclaw.bak.YYYYMMDD/workspace/SOUL.md ~/.openclaw/workspace/SOUL.md

# 3. If no backup: use Claude Code as external reviewer
# Claude Code can read files without being affected by OpenClaw's context
claude  # open Claude Code, read files, diagnose

Sandbox mode: never enable

On macOS with external drives + symlinks, sandbox mode kills the system:

Sandbox ON + symlinks to /Volumes/2TB_APFS/ = agent loses access to workspace

OpenClaw's sandbox mounts a container that can't follow symlinks to external volumes. Keep sandbox OFF always.

---

5. Shared memory architecture

Both Claude Code and OpenClaw read/write the same memory. No idea gets lost.

Obsidian vault (structural)          MemPalace (semantic / ChromaDB)
/obsidian-memory/                    /memory/palace/
  wiki/     ← docs, MOCs               ↑
  daily/    ← session logs             └─ mined from vault
  ideas/    ← captured ideas           
  answers/  ← resolved Q&A            Search: mempalace search "keyword"

Capture ideas from any agent

capture-idea \
  --title "OAuth tokens expire too fast on mobile" \
  --topic bezpieczenstwo \
  --tags "oauth,mobile,token" \
  --body "Tokens expire after 1h, users get logged out mid-session." \
  --source openclaw  # or: claude-code | lmstudio | manual

Mine vault into MemPalace

mempalace mine /path/to/obsidian-memory/ --wing obsidian_memory

Python 3.14 breaks MemPalace

MemPalace uses chromadb + pydantic v1. Python 3.14 breaks this. Fix: run MemPalace in a Python 3.12 venv.

brew install python@3.12
/opt/homebrew/bin/python3.12 -m venv /path/to/mempalace-venv
/path/to/mempalace-venv/bin/pip install mempalace

Wrapper at /opt/homebrew/bin/mempalace:

#!/bin/zsh
VENV=/path/to/mempalace-venv
PALACE=/path/to/palace
if [[ "$*" != *"--palace"* ]]; then
  exec "$VENV/bin/mempalace" --palace "$PALACE" "$@"
else
  exec "$VENV/bin/mempalace" "$@"
fi

---

6. Inter-agent communication: Claude Code ↔ OpenClaw

Claude Code can send messages to a running OpenClaw instance:

openclaw agent --message "Review this SQL migration — is it safe?" --agent main --json

OpenClaw responds, Claude Code parses the JSON response.

This enables:

• Claude Code asking Rook for a second opinion
• Rook asking Claude Code about security concerns (GRAY_ZONE consultations)
• Handoff between sessions

See: openclaw-bridge skill for full documentation.

---

7. WhatsApp console loop

OpenClaw has a WhatsApp gateway that creates a bidirectional loop:

• What you type in the console → appears in WhatsApp
• What you send from WhatsApp → appears in the console

This means nothing escapes your attention — if something happens at 3am while OpenClaw is running crons, you see it in WhatsApp in the morning.

Setup requires:

1. OpenClaw channel configured for WhatsApp
2. tts-whatsapp skill for voice output (optional)
3. Phone number linked in openclaw.json channels config

Console monitoring: openclaw logs --tail or watch the TUI.

---

8. Decision framework: when to act vs ask vs block

Every operation falls into one of three categories:

Category	Action	Claude offline?
SAFE	Do it	Do it
GRAY_ZONE	Ask Claude first	Defer 1h, retry
NEVER	Ask Claude first	Block, escalate to human

SAFE examples: generowanie, subagenci LM Studio, odczyt plików, web search na zaufanych domenach, capture-idea do ideas/

GRAY_ZONE examples: nowe crony, pip install --user, web fetch z nieznanych domen, modyfikacje ~/.config/

NEVER examples: edycja SOUL.md/AGENTS.md, rm -rf, instalacja modeli >35B, otwieranie portów, sandbox ON

Consult Claude via: openclaw agent --message '{"question":"...","risk":{...}}' --agent main --json

---

9. Backup strategy

# Before any major change:
openclaw backup create

# Manual snapshot:
cp -r ~/.openclaw ~/.openclaw.snap.$(date +%Y%m%d-%H%M)

# What to back up:
# ~/.openclaw/workspace/SOUL.md         ← identity
# ~/.openclaw/workspace/AGENTS.md       ← routing rules  
# ~/.openclaw/openclaw.json             ← gateway config (no secrets in git!)
# /path/to/obsidian-memory/             ← shared memory vault

Git the workspace (without secrets):

cd ~/.openclaw/workspace
echo "*.env\ncredentials*\n*.key" > .gitignore
git init && git add . && git commit -m "backup: workspace snapshot"

---

10. Skills that work together

Skill	What it does	Publish status
openclaw-bridge	Claude Code ↔ OpenClaw messaging	github.com/nerua1/openclaw-bridge
shared-memory-stack	Memory architecture reference	github.com/nerua1/shared-memory-stack
ralph	Persistence loop to completion	github.com/nerua1/ralph
ralph-wiggum-loop	Generator→Critic→Fixer→Verifier	github.com/nerua1/ralph-wiggum-loop
capture-idea	Save ideas from any agent	/opt/homebrew/bin/capture-idea
safe-skill-install	Security audit before installing	ClawHub: npx clawhub inspect
proactive-agent	Self-improving agent architecture	ClawHub: halthelobster
publish-skill	Publish skills to GitHub	Local: openclaw workspace/skills/

---

Quick health check

# Is OpenClaw gateway up?
openclaw health

# Is MemPalace working?
mempalace status
mempalace search "test"

# Is LM Studio loaded?
curl -s <lm-studio-url>/v1/models | python3 -c "import json,sys; [print(m['id']) for m in json.load(sys.stdin)['data']]"

# Can Claude Code reach OpenClaw?
openclaw agent --message "ping" --agent main --json | python3 -c "import json,sys; d=json.load(sys.stdin); print('OK:', d['status'])"

# Check RAM pressure
python3 -c "import psutil; m=psutil.virtual_memory(); print(f'RAM: {m.used/1e9:.1f}/{m.total/1e9:.1f}GB ({m.percent}%)')"

---

If this saved you time: ☕ PayPal.me/nerudek GitHub: github.com/nerua1