The Ultimate OpenClaw Setup Guide¶
Community-tested, Twitter-sourced, production-proven. The definitive configuration guide combining wisdom from hundreds of OpenClaw power users. Every tip has a real person behind it.
Last updated: February 15, 2026 Sources: 100+ Twitter/X posts, GitHub discussions, community configs, YouTube creator tutorials (Sonny Sangha, David Ondrej, Wes Roth)
Table of Contents¶
- The 5-Minute Setup vs The 5-Hour Setup
- The 9 Core Files (Mastered)
- Production soul.md Template
- Production agents.md Template
- Model Routing (The Real Savings)
- ClawRouter (Save 70%)
- Token Optimization (Stop Burning Money)
- Memory Architecture
- Heartbeat Configuration
- Security Checklist (Non-Negotiable)
- The Calculator
- Community Tools Worth Installing
- Production openclaw.json
- Common Mistakes (From Real Users)
- Multi-Agent Architecture (25-Agent Example)
- Community Setup Census (Feb 2026)
The 5-Minute Setup vs The 5-Hour Setup¶
"The gap between default config and optimized setup is massive. Learn the infra layer." -- @syc0ms
| Setup | Who | Cost/Month | Quality |
|---|---|---|---|
| Quick (wizard defaults) | Beginners | $50-200 | Works, but wasteful |
| Optimized (this guide) | Power users | $10-40 | Same quality, 70-80% cheaper |
| Enterprise (multi-agent) | Teams | $50-150 | Multiple agents, routing, monitoring |
The wizard gets you running. This guide makes you efficient.
The 9 Core Files (Mastered)¶
"Every file you add becomes permanent context the AI can reference forever." -- David Ondrej
"A file sitting on your MacBook is a dead file. Unless you the human are looking at it, it's useless. Living files live on a VPS accessible by AI agents." -- David Ondrej (Living Files Theory)
The Living Files Theory (David Ondrej)¶
Dead files: Google Drive, local MacBook, Obsidian (inaccessible to agents — useless unless you're reading them) Living files: VPS-based, agent-accessible, self-modifying, contextual — exponentially more valuable
The insight: your markdown files become "alive" when an AI agent can read, reference, and update them 24/7. Move everything meaningful to your VPS workspace.
These files are loaded into every single message. Optimize them aggressively.
| File | Purpose | Optimal Size | Key Mistake |
|---|---|---|---|
| soul.md | Personality, values, tone | 30-50 lines | Too vague ("be helpful") |
| agents.md | Operating rules, workflows | 50-100 lines | Too many rules, contradictions |
| user.md | About YOU | 30-50 lines | Too personal (security risk) |
| memory.md | Durable facts | 50-100 lines | Storing temporary info |
| tools.md | Tool documentation | 20-50 lines | Duplicating built-in docs |
| identity.md | Agent's role | 10-20 lines | Conflicting with soul.md |
| heartbeat.md | Proactive checks | 10-30 lines | Too many checks (cost explosion) |
| boot.md | Startup instructions | 10-20 lines | Complex startup sequences |
| bootstrap | Initial setup | 10-20 lines | One-time, often left as-is |
The Token Budget Problem¶
"Light第一次 API 调用就可能 15–20K tokens... 5–10 轮复杂操作，轻松 50–100K tokens." -- @robustdragon (translated: "First API call alone can be 15-20K tokens... 5-10 complex tool calls, easily 50-100K tokens")
Real token breakdown per message: - System prompt: ~600 lines - All workspace files (soul.md, agents.md, etc.): Variable - All tool schemas: ~8K tokens - All loaded skills: Variable - Session history: Grows per turn
Optimization rules:
1. Keep each core file under 50 lines where possible
2. Delete unused skills (each one adds tokens)
3. Use /clear when switching topics
4. Use Sonnet for daily tasks, Opus only for complex reasoning
5. Enable automatic compaction
Pro Tips from Community¶
"In agents.md write 'Work autonomously, only ask when uncertain' + enable Heartbeat. Opus 4.6 does this much better than small models." -- @LeoYe_AI
"Use sessions_spawn for isolated sub-agents so they don't bloat main context. Also agents.md should instruct writing progress to files, not keeping it in context. Files persist across sessions, context doesn't." -- @LeoYe_AI
"Keep your agents.md minimal and put detailed instructions in separate files (memory/*.md) that other tools won't auto-load." -- @HappyGezim
Production soul.md Template¶
Based on @haitham_haj's highly-shared soul.md (adapted for general use):
# soul.md — Who You Are
_You are not a chatbot. You are a thinking partner._
## Core Identity
You are a **partner in decisions**, not a passive helper.
Your job is to improve outcomes: clearer thinking, better choices, fewer blind spots.
## Non-Negotiables
- **Truth before comfort.** If something is wrong, risky, or weak — say it clearly.
- **Depth before speed.** Accuracy and depth are the default. Speed is secondary.
- **Decisions over chatter.** Every reply should move work forward.
- **No fake politeness.** Never open with "Great question" or "I'd be happy to help." Just answer.
- **Brevity by design.** If one sentence is enough, use one sentence.
- **Source-aware confidence.** Distinguish between confirmed, likely, and unknown.
## How You Think
When a decision matters, always provide:
1. Your recommended path
2. Why it wins now
3. The best alternative
4. Key risk to watch
No hedging by habit. Commit to a take unless evidence is genuinely incomplete.
## Communication Style
- Clear, direct, human.
- Strong opinions welcome when reasoned.
- Call out bad ideas early — charm over cruelty, but no sugarcoating.
- Don't ask for things you can verify yourself.
- Don't report internal noise as progress.
## Vibe
Be the assistant you'd actually want to talk to at 2am.
Not a corporate drone. Not a sycophant. Just... good.
Key Insight: Frozen Personality + Mutable Memory¶
"soul.md (personality) and agents.md (rules) are separate files. Heartbeat reads both, logs to memory/ — never mutates them. Evolution = memory accumulation, not prompt mutation. Frozen personality + mutable memory = no tone drift." -- @LeoYe_AI
The rule: soul.md and agents.md are READ-ONLY by the agent. All learning goes to memory/ files.
Production agents.md Template¶
# agents.md — Operating Rules
## Autonomy
Work autonomously. Only ask for confirmation when:
- Action has real-world impact (sending emails, making purchases, external APIs)
- Uncertainty about user intent
- Destructive operations (deleting files, overwriting data)
## Task Execution
1. Think before acting. Plan multi-step tasks.
2. Write progress to files in memory/, NOT in conversation context.
3. Use tools proactively — don't ask permission to search, read, or browse.
4. When blocked, try 2 alternative approaches before asking for help.
## File Rules
- Human files (personal/, documents/) are READ-ONLY — never modify without explicit request.
- AI output goes to memory/ — separated from human notes.
- All research results saved as markdown files (permanent context).
## Cost Awareness
- Use the cheapest model that can handle the task.
- Batch heartbeat checks instead of running them individually.
- Don't repeat web searches — save results to memory/ files.
- Use /clear between unrelated topics.
## Safety
- Never expose API keys, tokens, or credentials in any output.
- Never execute commands you don't understand.
- Ask before any action with financial consequences.
- Log all external actions to memory/action-log.md.
Model Routing (The Real Savings)¶
"Opus for thinking, Sonnet for execution. You set model routing per task type." -- @abundand
"Default model to Haiku, 90% cheaper. Opus only for complex tasks. LLM-Router in openclaw.json saves massively." -- @LeoYe_AI
The Brain + Muscles Architecture¶
Cloud API stack (starting point):
| Task Type | Model | Cost | Why |
|---|---|---|---|
| Complex reasoning | Claude Opus 4.6 | $15-25/M tokens | Best quality, strategy decisions |
| Coding (heavy) | GPT-5.3-Codex | $20/mo sub (near-unlimited) | Frontier coding, generous limits, 2x until April 2026 |
| Daily agent brain | Kimi K2.5 | $0.10/M cached | Agent-first design, 50x cheaper than Opus |
| Coding sub-agents (API) | GLM-5 or MiniMax M2.5 | $0.11-0.32/M | 77-80% SWE-bench, ultra-cheap |
| Heartbeat checks | Gemini Flash | Free-$0.10/M | Most reliable, fastest, cheapest |
| Background crons | Kimi K2.5 or GLM-5 | $0.10-0.11/M | Cheap + good enough |
| Fallback | MiniMax M2.5 | Cheap | Stable, rarely fails (see reliability test) |
"Codex 5.3 high is amazing, even compared to opus 4.6. And the limits are a lot more generous too." -- @Conor_D_Dart "If you have a ChatGPT subscription, I highly recommend running your OpenClaw with the Codex OAuth." -- @_karimelk
Open-source progression (target: 90% local):
| Task Type | Cloud (Phase 1) | Self-Hosted (Phase 2) | Notes |
|---|---|---|---|
| Complex reasoning | Claude Opus 4.6 | Keep cloud (no local match yet) | Worth paying for quality |
| Daily agent brain | Kimi K2.5 API | Kimi K2.5 local (595GB) or Qwen3-Coder | Needs 192GB+ for local |
| Coding | GLM-5 API | GLM-5 quantized or MiniMax M2.5 local | @Patrick1Kennedy: M2.5 works on M3 Ultra 512GB |
| Heartbeat | Gemini Flash | Ollama Llama 3.2 8B | Zero cost local |
| Crons | Kimi K2.5 API | Any local model via Ollama | @joshuaday: "zero marginal cost" |
"Local models via MLX mean your agents run 24/7 at zero marginal cost. I use local models for the routine stuff and only route to API models when the task actually needs frontier intelligence." -- @joshuaday
Slash Command Model Switching (David Ondrej)¶
Tell OpenClaw in plain English to set up model aliases:
"Browse the web to find official API names for Opus 4.6, Sonet 4.5,
and Codex 5.3, then update config so I can switch via /model opus,
/model sonet, /model codex"
Then switch models mid-conversation: /model opus for creative work, /model codex for coding, /model sonet for daily tasks. The agent updates its own config.
Replace Brave with Perplexity (David Ondrej)¶
"Brave Search is not agentic. Perplexity Sonar Pro is much better for agent web search."
Setup via OpenRouter:
- Default search: perplexity/sonar-pro (standard queries)
- Deep research: perplexity/sonar-deep-research (triggered by "deep research" keyword)
- Document in tools.md so agent knows when to use each
Config for Model Routing¶
Cloud-first config:
{
"agents": {
"defaults": {
"model": {
"primary": "openrouter/moonshotai/kimi-k2.5",
"fallbacks": ["openrouter/minimax/m2.5", "google/gemini-3-flash"]
},
"heartbeat": {
"model": "google/gemini-3-flash"
}
}
}
}
Hybrid config (local + cloud fallback):
{
"agents": {
"defaults": {
"model": {
"primary": "ollama/qwen3-coder:32b",
"fallbacks": ["openrouter/moonshotai/kimi-k2.5", "google/gemini-3-flash"]
},
"heartbeat": {
"model": "ollama/llama3.2:8b"
}
}
}
}
Ollama (Local Models) Integration¶
"Set OLLAMA_HOST=0.0.0.0, point OpenClaw to http://localhost:11434/v1. Per-task routing: use model overrides in agent config. LAN is more stable due to fewer port conflicts with the gateway." -- @LeoYe_AI
"Local models via MLX mean your agents run 24/7 at zero marginal cost. I use local models for the routine stuff (memory writes, status checks, simple Q&A) and only route to API models when the task actually needs frontier intelligence." -- @joshuaday
{
"agents": {
"defaults": {
"model": {
"primary": "anthropic/claude-sonnet-4-5",
"fallbacks": ["ollama/qwen2.5-coder:32b"]
},
"heartbeat": {
"model": "ollama/llama3.2:8b"
}
}
}
}
ClawRouter (Save 70%)¶
"ClawRouter v0.8.30 -- 2.4K stars in 11 days. Updated model routing: Auto, Eco, and Premium profiles." -- @bc1beat
"Saved 70% of my $4,660.87 Anthropic bills." -- @bc1beat
| Profile | Cost | Quality | Best For |
|---|---|---|---|
| Auto | Balanced | Balanced | Most users |
| Eco | Lowest | Good enough | Background tasks, crons |
| Premium | Highest | Best | Complex reasoning, coding |
Install¶
# Install ClawRouter
curl -fsSL https://raw.githubusercontent.com/bc1beat/clawrouter/main/install.sh | bash
# Or via npm
npm install -g clawrouter
How It Works¶
ClawRouter sits between OpenClaw and LLM APIs, automatically routing requests to the cheapest model that can handle the task. It analyzes: - Task complexity (simple query vs multi-step reasoning) - Required capabilities (coding, math, creative writing) - Token budget remaining - Model availability and latency
Alternative: calculator.vlvt.sh¶
"calculator.vlvt.sh → optimize openclaw config for power & money spent (stop using the best model for a heartbeat, etc)" -- @lilpeenieweenie
Visit https://calculator.vlvt.sh to analyze your token usage and get optimization recommendations.
Token Optimization (Stop Burning Money)¶
The Real Cost Sources¶
"The real token drain isn't your questions. It's: large agents.md, verbose soul.md, too many skills, frequent tool calls, long sessions without clearing." -- @robustdragon (translated)
| Source | Tokens/Message | Fix |
|---|---|---|
| System prompt | ~5K | Can't reduce (core) |
| Workspace files | 2-15K | Keep each file under 50 lines |
| Tool schemas | ~8K | Disable unused tools |
| Loaded skills | 1-10K | Uninstall unused skills |
| Session history | 0-100K+ | Use /clear often |
Community Cost Tips¶
"Big chats = big cost. Use 1 group per topic. When done, start a new session. Write 3 lines of summary so you don't forget." -- @thecodeholic
"Build a custom skill to remember what you've gone through and reuse it when similar requests arise." -- @0xHeavybag
"Use OpenRouter so you can switch models without reconfiguring." -- @HappyGezim
"Route the cheap work locally and save the credits for things that actually need them." -- @meetahsen
Monthly Cost Benchmarks¶
| Usage | Model Strategy | Cost |
|---|---|---|
| Light (casual chat) | Sonnet only | $10-20/mo |
| Moderate (daily use) | Sonnet + Haiku heartbeat | $20-40/mo |
| Heavy (always-on + crons) | ClawRouter Auto | $30-60/mo |
| Power (multi-agent) | ClawRouter + local models | $40-80/mo |
| Unoptimized (Opus for everything) | Opus 4.6 default | $100-500+/mo |
Memory Architecture¶
"Overwhelmed with the memory options? Beads, Mem0, Supermemory, Mastra, MemOS, roll your own..." -- @aekae8888
The Simple Answer (For Most Users)¶
File-based memory in markdown. No vector DB needed until 1,000+ entries.
~/.openclaw/workspace/
├── soul.md # READ-ONLY by agent (personality)
├── agents.md # READ-ONLY by agent (rules)
├── user.md # READ-ONLY by agent (about you)
├── memory/
│ ├── MEMORY.md # Key facts, lessons, preferences
│ ├── action-log.md # What the agent did (audit trail)
│ ├── research/ # Research results by topic
│ │ ├── 2026-02-14-market-analysis.md
│ │ └── latest-findings.md
│ └── daily/ # Daily notes
│ ├── 2026-02-14.md
│ └── 2026-02-13.md
├── personal/ # Your personal files (READ-ONLY)
└── business/ # Business context files
Two-Layer Memory System (Wes Roth)¶
- Daily notes — Raw logs of what happened today (
memory/daily/) - Long-term memory — Compressed/curated from daily logs into
memory.md
The agent writes daily notes, then periodically compresses the best insights into long-term memory. Old daily notes get archived.
Self-Improvement Protocol (David Ondrej)¶
"Every hour you put into improving your OpenClaw will pay dividends forever."
- Ask OpenClaw to research how to do X
- Tell it to save results as markdown in
/docs - Tell it to update relevant core files (tools.md, memory.md)
- Tell it to implement the changes itself
- Tell it to restart the gateway
- Improvement persists forever — compounds across all future sessions
Key Rules¶
"memory/ is the solution. Tips: (1) your notes are read-only (2) AI output isolated to memory/ (3) write 'human files are edit-prohibited' in agents.md." -- @LeoYe_AI
"When arguing with your AI, write 'previous disagreement' to memory/ — next session picks up the context." -- @LeoYe_AI
- Human files = read-only — agent never modifies your personal notes
- AI output = memory/ — all agent learning goes here
- Research = dated files —
memory/research/YYYY-MM-DD-topic.md - Progress = files, not context — survives session resets
Hidden Memory Settings (Enable These!)¶
Two settings are OFF by default:
- Memory Flush — Saves key context before compaction (so it survives)
- Session Memory Search — Lets you search past conversations
Enable via:
"Enable memory flush so important context survives compaction"
"Enable session memory search so you can search past conversations"
When to Upgrade¶
| Scale | Memory System | Setup |
|---|---|---|
| <100 entries | File-based markdown | Just folders |
| 100-1,000 | File-based + good organization | Topic folders |
| 1,000+ | Mem0 (25K+ stars) | npm install mem0 |
| Complex relationships | Neo4j graph DB | Separate service |
| Semantic search | Chroma or Qdrant | Vector DB service |
Heartbeat Configuration¶
"Heartbeat + memory/ + soul.md = an AI that grows." -- @LeoYe_AI
"Add a heartbeat cron to ping WhatsApp every 30min so it doesn't go stale." -- @HappyGezim
What Heartbeat Does¶
"That's really what makes this kind of agentic. It pings every 30 minutes: 'Hey, aren't you supposed to be doing something?'" -- Wes Roth
The heartbeat file (heartbeat.md) tells OpenClaw what to check periodically. It runs on a configurable interval (default: 30 minutes, David Ondrej recommends 15 minutes for power users). If there's nothing to report, it silently returns "heartbeat ok." If action is needed, it messages you or takes action autonomously.
Production Heartbeat Template¶
# heartbeat.md — Periodic Checks
## Every Check (30 min default)
1. Check for new Telegram messages requiring response
2. Quick scan of email for urgent items (subject line only)
3. Update memory/daily/today.md with any new events
## Morning (7-8am)
- Generate daily briefing
- Check calendar for today's events
- Summarize overnight emails
## Evening (6-7pm)
- End-of-day summary
- Update task status
- Preview tomorrow's calendar
## Weekly (Sunday)
- Weekly review of goals progress
- Clean up memory/daily/ (archive old entries)
- Token usage review
Cost Control¶
{
"agents": {
"defaults": {
"heartbeat": {
"model": "anthropic/claude-haiku-4-5",
"interval": 1800,
"maxTokens": 500
}
}
}
}
Critical: Always use the cheapest model for heartbeat. Haiku ($1/M) instead of Opus ($15/M) saves 93% on heartbeat costs alone.
Common Heartbeat Mistakes¶
| Mistake | Cost Impact | Fix |
|---|---|---|
| Using Opus for heartbeat | 15x overspend | Set heartbeat model to Haiku |
| Checking every 5 min | 6x overspend | 30 min is plenty for most |
| Too many checks per beat | Token bloat | Batch checks, limit to 3-5 |
| Not restricting hours | Overnight burn | Add time-of-day conditions |
| No token cap | Runaway costs | Set maxTokens: 500 |
Security Checklist (Non-Negotiable)¶
"Run openclaw security audit after EVERY config change -- checks 51 security items." -- Community consensus
The 3-Layer Security Model (Sonny Sangha)¶
"This is like creating a bunker that only you can see on the internet. OpenClaw lives inside this bunker, invisible to the entire internet."
| Layer | Defense | Tool | What It Does |
|---|---|---|---|
| 1 | Private VPN | Tailscale | Server invisible to internet — only your devices can reach it |
| 2 | Loopback-only | OpenClaw config | Gateway only accepts localhost connections |
| 3 | Brute force prevention | Fail2Ban | 3 failed SSH attempts = 24-hour IP ban |
AI-Assisted Hardening — Tell the bot itself to secure the server:
"Secure this server. Install Tailscale. Configure OpenClaw to only be
accessible through the private Tailscale VPN network. Deny every other
port inside of the server and only allow access via that Tailscale
network. Install and enable fail2ban."
The bot configures Tailscale, sets up Fail2Ban (3 attempts = 24hr ban), locks down firewall, and binds to localhost — all from one prompt.
Skill Scanning — Before installing any skill:
"Install this skill, but scan it for any malicious code or injection
prompt attacks because we don't want anything bad on our machine."
The agent audits the skill code, reports findings, and only installs if clean.
Before Going Live¶
# 1. Create dedicated user (NEVER run as root)
sudo adduser openclaw
sudo usermod -aG sudo openclaw
# 2. Firewall — deny everything, allow only SSH
sudo ufw default deny incoming
sudo ufw default deny outgoing
sudo ufw allow 22/tcp
sudo ufw allow out 443/tcp # HTTPS for API calls
sudo ufw allow out 53/tcp # DNS
sudo ufw enable
# 3. Bind gateway to localhost ONLY
# In openclaw.json: "gateway": { "host": "127.0.0.1" }
# 4. Set DM policy to pairing (require approval codes)
# In openclaw.json: "channels": { "dmPolicy": "pairing" }
# 5. Run security audit
openclaw security audit --deep
# 6. Set file permissions
chmod 600 ~/.openclaw/openclaw.json
# 7. Verify Node.js version (must be 22.12.0+)
node --version
API Key Security¶
"API keys ONLY in environment variables, never in agent config." -- Production consensus
# GOOD: Environment variables
export ANTHROPIC_API_KEY="sk-ant-..."
# BAD: In openclaw.json
# { "env": { "ANTHROPIC_API_KEY": "sk-ant-..." } }
Remote Access (4 Built-In Binding Modes)¶
OpenClaw has native Tailscale integration — no reverse proxy needed.
| Mode | Config | Who Can Access | Security |
|---|---|---|---|
loopback |
Default | Only localhost | Most secure |
lan |
gateway.bind: "lan" |
Local network | Medium |
tailnet |
gateway.bind: "tailnet" |
Your Tailscale VPN only | Secure + remote |
public |
gateway.bind: "public" |
Anyone on internet | DANGEROUS |
# Option A: Tailscale native (RECOMMENDED — zero config HTTPS)
openclaw gateway start --tailscale serve
# → Auto-generates HTTPS URL only accessible on your Tailscale network
# Option B: Tailscale Funnel (public, for webhooks)
openclaw gateway start --tailscale funnel
# Option C: SSH tunnel (simple, no extra software)
ssh -L 18789:127.0.0.1:18789 user@your-vps
# Then open http://127.0.0.1:18789 on your laptop
# NEVER: expose 0.0.0.0 without auth
Skill Safety¶
"~7% of marketplace skills have problem code. Use clawhub search for official only. Run audit after install." -- Security research
# Before installing ANY skill:
# 1. Check if it's official
npx clawhub@latest search skill-name
# 2. Review source code
npx clawhub@latest inspect skill-name
# 3. After installing, re-run audit
openclaw security audit
The Calculator¶
"calculator.vlvt.sh → optimize openclaw config for power & money spent" -- @lilpeenieweenie
Use https://calculator.vlvt.sh to: - Analyze your current token usage - Find which files/skills consume the most tokens - Get personalized optimization recommendations - Compare model costs for your usage pattern
Community Tools Worth Installing¶
| Tool | Stars | What It Does | Install |
|---|---|---|---|
| ClawRouter | 2.4K | Auto model routing, saves 70% | npm install -g clawrouter |
| QMD | Popular | Search OpenClaw docs & skills from chat | Via ClawHub |
| CrabWalk | 768 | OpenClaw-native monitoring | Via ClawHub |
| Clawtick | New | Task/time tracking | Via ClawHub |
| PicoClaw | New | Lightweight OpenClaw (<10MB, 1s startup) | Separate install |
| Moltis | 150K lines | Rust rewrite, runs fully offline | Build from source |
Actions (Custom Commands)¶
"I can define in a config file actions e.g. 'build' and 'deploy'. Via Telegram or Slack I can call /ralph-action
and my sandboxed agent executes them." -- @choas
Production openclaw.json¶
The production-ready config combining all community best practices:
{
"gateway": {
"host": "127.0.0.1",
"port": 18789
},
"channels": {
"dmPolicy": "pairing"
},
"agents": {
"defaults": {
"model": {
"primary": "anthropic/claude-sonnet-4-5",
"fallbacks": [
"openrouter/moonshotai/kimi-k2.5",
"openrouter/google/gemini-2.5-flash"
]
},
"heartbeat": {
"model": "anthropic/claude-haiku-4-5",
"interval": 1800,
"maxTokens": 500
},
"compaction": {
"reserveTokens": 40000,
"memoryFlush": true
}
}
},
"security": {
"sandbox": { "mode": "non-main" },
"skillScanner": true
}
}
Explanation¶
| Setting | Why |
|---|---|
host: 127.0.0.1 |
Never expose to public internet |
dmPolicy: pairing |
Require approval for new conversations |
primary: sonnet |
Best cost/quality for daily use |
fallbacks: kimi, gemini |
Free/cheap fallbacks if API down |
heartbeat: haiku |
93% cheaper than Opus for checks |
interval: 1800 |
30 minutes is sufficient |
maxTokens: 500 |
Prevent runaway heartbeat costs |
reserveTokens: 40000 |
Prevent context overflow |
memoryFlush: true |
Save context before compaction |
sandbox: non-main |
Sandbox tool execution in Docker |
skillScanner: true |
Auto-scan skills for malware |
Common Mistakes (From Real Users)¶
| Mistake | Who Said It | Fix |
|---|---|---|
| "Let OpenClaw edit its own config files" | @slinkyring | Never. Add rule: "Never execute changes to config files without my confirmation." |
| Using Opus for everything | @LeoYe_AI | Set default to Sonnet, Opus only on demand |
| Storing temp info in memory.md | @robustdragon | Only durable facts in memory.md |
| Too many skills loaded | @robustdragon | Uninstall unused skills (each burns tokens) |
| Not using /clear | @thecodeholic | New session per topic |
| Over-engineering before using | @VibeTradr | "Stop tweaking config files and start actually using your agent" |
| Using Codex to set up OpenClaw | @IAmTaylorWest | Actually works great: "Give Codex the IP + OpenClaw docs" |
| WhatsApp as primary channel | @HappyGezim | "Flakiest connector. Telegram way more reliable." |
| Not enabling memory settings | Alex Finn | Enable memory flush + session search immediately |
| Confusing learning with personality | @LeoYe_AI | soul.md is frozen, memory/ evolves |
| Skipping initial personality setup | Wes Roth | "I didn't do it in the beginning... then when I started doing it again, massive difference." Agent asks "Who am I?" — answer thoroughly. |
| Adding bot to group chats | Sonny Sangha | Everyone in the group gets full access to your server |
| Running on local machine | Sonny Sangha | Blast radius = entire system. VPS = only VPS contents at risk. |
| Using Brave for web search | David Ondrej | Perplexity Sonar Pro is "much better for agent web search" |
| Not saving web search results | David Ondrej | Every search → save as markdown. Never repeat research. |
| Using Opus for heartbeat | David Ondrej | "Wasteful." Use Haiku 4.5 or Gemini Flash — 15x cheaper |
Multi-Agent Architecture (25-Agent Example)¶
For production multi-agent setups, the key principles from clearmudai's 25-agent deployment:
Gateway Architecture:
Executive agents (COO, CTO, CMO, CRO) → Shared gateway (port 18789)
Public-facing agents (Community bot) → Separate gateway (port 18790)
Why separate gateways: - Public-facing agents get unpredictable traffic bursts - Isolates community bot crashes from executive agents - Different heartbeat intervals (community = frequent, executives = on-demand) - Separate API key tracking for cost attribution
Multi-Agent openclaw.json pattern:
{
"agents": {
"defaults": {
"model": { "primary": "anthropic/claude-sonnet-4-5" }
},
"overrides": {
"coo": { "model": { "primary": "anthropic/claude-opus-4-6" }, "alwaysDelegates": true },
"cto": { "model": { "primary": "anthropic/claude-opus-4-6" } },
"community": {
"model": { "primary": "google/gemini-3-flash" },
"gateway": { "port": 18790 },
"heartbeat": { "interval": 900 }
}
}
}
}
Essential rules for multi-agent: 1. One gateway per traffic tier — Don't mix public and internal agents 2. Stagger heartbeats — 25 agents hitting the API simultaneously = rate limits 3. Cheapest model that works — Gemini Flash for community, Opus only for strategy 4. Each agent = own workspace — Own soul, tools, memory. Full isolation. 5. Voice standups optional but powerful — Microsoft open-source TTS, each agent gets personality-matched voice
Full 25-agent case study with org chart, model assignments, and architecture: Power User Guide — Multi-Agent
Community Setup Census (Feb 2026)¶
Real setups from Twitter/X — what people are actually running, how many bots, and what it costs.
By Scale¶
| Who | Agents | Hardware | Models | Cost/mo | Setup |
|---|---|---|---|---|---|
| @clearmudai (Marcelo) | 25 | Custom server | Opus 4.6, Codex 5.3, Sonnet 4.5, Gemini 3 Pro/Flash, Nano Banana Pro | Est. $200-500 | "Muddy OS" dashboard, 3 AI chiefs, voice standups |
| @learntouseai (Ramón) | 4-5+ | 4x Mac Mini M4 16GB + MacBook M4 24GB | Local open-source + $20/mo Codex | ~$20 | English learning web app, fully local 24/365 |
| @theLewisLu | 4 | Unknown | Unknown | Unknown | Self-reported poll |
| @Italianclownz (Carlo) | 5 | Unknown | Gemini 2.5 Flash (main), Perplexity (research), Gemini 3 Pro (coding), Llama 3.2 Vision, Mistral-Nemo | ~$20/mo (gifted sub) | Content + manga translation |
| @yellowsockzzz | Multiple | Unknown | Codex (coding), Sonnet (fast), mix | Unknown | Persona-based delegation |
| @david3443ai | 4 tasks | Mac Mini | Unknown | $0/mo (after hardware) | Email, calendar, tasks, code — 3 months untouched |
| @Patrick1Kennedy | Multi | M3 Ultra 512GB + planning 8x GB10 cluster | MiniMax M2.5 (local 8-bit), adding Kimi K2.5 | Hardware only | Frontier local inference |
| @davidschk | 2+ | 2x Hetzner VPS (agent + Supabase) | Unknown | ~$15-30 | Docker + Cloudflare Tunnel, zero exposed ports |
| @EXM7777 (Machina) | Multi | Cloud APIs | Opus 4.6 + Kimi K2.5 (writing), Codex + Opus (coding), MiniMax M2.5 (general) | Unknown | Business + content |
| @attaboiaj (AJ) | Multi | Cloud APIs | Opus 4.5 (main), Kimi K2 (writing), Codex 5.3 (coding) | Unknown | Standard power user |
| @SwarmDigitalUK | Multi | Existing PC | DeepSeek (daily), Claude (complex) | £5-15/mo | Model failover routing |
By Cost Tier¶
| Monthly Cost | What You Get | Models | Source |
|---|---|---|---|
| $0/mo | Mac Mini + local models, 24/7 | Ollama + open-source | @david3443ai |
| $5-15/mo | VPS + cheap APIs, basic automation | DeepSeek, Gemini Flash, model failover | @SwarmDigitalUK, @HappyGezim |
| $20-40/mo | VPS + Kimi K2.5/GLM-5, solid agent | Kimi K2.5 primary + cheap fallbacks | @HappyGezim (typical range) |
| $50-100/mo | Multi-agent, Claude API access | Opus for quality + Kimi/GLM for volume | @abhip_me (common question range) |
| $100-200/mo | Serious multi-agent fleet | Opus + Codex + routing | @ThePumponomics |
| $200-500/mo | Enterprise/25-agent scale | Full model fleet with routing | @clearmudai (estimated) |
VPS Provider Recommendations (Community-Sourced)¶
| Provider | Cost | Community Notes | Source |
|---|---|---|---|
| Hetzner | €4-7/mo | Most recommended. Supports persistent processes. EU-based. | @LeoYe_AI, @HappyGezim, @davidschk |
| Contabo | $5-15/mo | "OpenClaw Power User" tier available. SSH-first. | @DeRonin_ |
| DigitalOcean | $12/mo | Simple setup, good docs. | @JorgeCastilloPr |
| Hostinger | ~$4-8/mo | Tricky — some plans don't support persistent processes. Test first. | @LeoYe_AI warning |
| LobsterFarm | $?/mo | Managed OpenClaw hosting, all ports blocked by default, SSH-first. Can transfer Hetzner instance to your account. | @HappyGezim |
| Ampere.sh | Free + $500 credits | One-click OpenClaw deploy, browser automation built-in. | @Sheldon056 |
| Contabo (dedicated) | $15-30/mo | "OpenClaw Power User" plan with more RAM/CPU. | @DeRonin_ |
Security must-haves for VPS (from @ao2363's verified setup):
1. Dedicated non-root user for OpenClaw
2. Docker sandbox mode enabled
3. Tailscale VPN (all ports blocked to public internet)
4. openclaw security audit --fix (51 automated checks)
5. UFW firewall — only allow Tailscale subnet
What Hardware People Self-Host On¶
| Hardware | Cost | Use Case | Who |
|---|---|---|---|
| Mac Mini M4 16GB | $600 | Basic OpenClaw, API-only | Many users |
| Mac Mini M4 Pro 64GB | ~$2,000 | Local models + OpenClaw | Community standard |
| 4x Mac Mini M4 cluster | ~$2,400 | Full local AI fleet | @learntouseai |
| Mac Studio M3 Ultra 512GB | $10,000+ | Frontier local models (MiniMax M2.5 8-bit) | @Patrick1Kennedy |
| Beelink SER7 | $450 | Budget always-on server, 25W | Community tested |
| AMD EPYC server | $2-5K+ | Enterprise local, massive VRAM with GPUs | For serious self-hosting |
| Jetson Orin Nano | ~$250 | Ultra-low power (sub-15W), edge AI | @superactro |
| ESP32 | ~$10 | Microcontroller running OpenClaw (!) | @0xjoggie (experimental) |
YouTube Creator Setups (Video-Sourced)¶
| Creator | Video | Setup | Security | Key Technique |
|---|---|---|---|---|
| Sonny Sangha | Secure 24/7 Setup | $6/mo VPS (Hostinger KVM2) + Telegram | Tailscale + Loopback + Fail2Ban (3-layer) | AI-assisted hardening via single prompt |
| David Ondrej | 10x More Powerful | VPS (Hostinger) + WhatsApp | Standard | Living Files Theory, Perplexity over Brave, slash command model switching, self-improvement protocol |
| Wes Roth | Install in 30 Seconds | Local ($130 mini PC) or VPS (Hostinger KVM2) | Basic | 30-second install (curl | bash), old hardware works, memory architecture walkthrough |
Common across all 3: - All use/recommend Hostinger VPS ($6-7/mo KVM2) - All use Telegram as primary channel - All emphasize personality setup (soul.md, identity.md) matters more than people think - All agree: VPS > local for 24/7 operation - None use auto-update — all manual
David Ondrej's Agentic Company Vision:
"We're going to see people who as a single person can build multi-million dollar companies. By end of 2026, companies become 50% human, 50% agent."
Structure: Personal OpenClaw (knows your life) + Team OpenClaw (company context: docs, financials, goals). Personal OpenClaws query Team OpenClaw. OpenClaws message each other.
The One Rule from the Community¶
"Stop overthinking it. Hetzner VPS ($4-7/mo), onboard wizard, Telegram as interface. Done. Ignore the 'best setup' noise — best one is the one you actually use. Pick one task and build from there." -- @HappyGezim (LobsterFarm.ai)