Set up your agent right.

Answer a few questions. Get a complete config with the right models, cost-saving tips, and capability manifest — ready to paste into OpenClaw.

1OS
2Packs
3Provider
4Budget
5Channel
6Heartbeat
7Config
What's your operating system?
This changes the install commands, background service, and setup steps throughout the wizard.
✓ 🍎
macOS
Homebrew + launchd daemon. Apple Silicon or Intel.
Easiest
✓ 🐧
Linux
apt + systemd. Ubuntu 22.04+ recommended.
Easy
✓ 🪟
Windows
Requires WSL2 (Windows Subsystem for Linux). Not native.
WSL2 Required
What should your agent do?
Select one or more agent packs. Each adds verified skills to your setup guide. Pick as many as you need — we'll deduplicate the install commands.
Include Cross-Pack Essentials — 6 high-download skills every agent benefits from (Gog, Tavily, Summarize, ClawdHub, Skill Vetter, Docs)
Choose your model provider
This is where your API key comes from. Already have an API key? Pick that provider. Not sure? Use the recommendation.
🧭
Recommend for me
Not sure which provider to pick? We'll recommend one based on community consensus.
Or choose manually:
✓ Community #1 🟠
Anthropic
Claude Sonnet 4.5 — strong tool calling and SKILL.md compliance. Premium pricing ($3/$15 per 1M tokens).
From $1/M tokens
→ console.anthropic.com/keys
✓ 🟢
OpenAI
GPT-5.2, GPT-5.1, GPT-4o-mini. Strong ecosystem, competitive pricing ($1.25/$5 per 1M for GPT-5.1).
From $0.10/M tokens
→ platform.openai.com/api-keys
✓ 🔵
Google
Gemini 2.5 Pro and Flash. Excellent price-performance, especially Flash at $0.15/M.
From $0.08/M tokens
→ aistudio.google.com/apikey
✓ 🔀
OpenRouter
One API key, all models. Claude, GPT, Gemini, Llama, Mistral — mix and match freely.
Varies by model
→ openrouter.ai/keys
✓ 🦙
Ollama (Local)
Run models on your hardware. Qwen3, Llama, Mistral. Zero API cost, full privacy.
$0 — hardware only16GB+ RAM
→ ollama.com/download
Set your model tier
This controls which models your agent uses. The cost estimates update live as you move the slider — based on real token pricing, not guesses.
~$25/mo
Model tier — actual cost depends on usage
$0~$10~$25~$50~$75$100+
Models at this tier
How will you talk to your agent?
Pick your primary channel. You can add more later — all channels can run simultaneously.
✓ Recommended ✈️
Telegram
Easiest to set up. Official Bot API — zero ban risk. Works behind NAT, no inbound ports needed.
Setup: @BotFather → /newbot → copy token → paste → done in 5 minutes
✓ 💬
WhatsApp
QR code scan — uses your existing account via Linked Devices. Popular if WhatsApp is your daily driver.
⚠ ️ Caveat: Unofficial library. Consider a separate eSIM ($5-10/mo) to protect your main number.
✓ 🎮
Discord
Create a bot at discord.com/developers. Good for teams and developer workflows.
Setup: Create app → enable Message Content Intent → get bot token → invite to server
✓ 💼
Slack
Workspace app — ideal for work contexts and team-shared agents.
Setup: Create Slack app → Socket Mode → install to workspace → get bot + app tokens
Configure your agent's heartbeat
Your agent wakes up every 30 minutes to check on things. Pick what it should monitor — you can always edit this later by asking your agent to update HEARTBEAT.md.

✨ Want something more specific?

▼
Describe what you'd like your agent to monitor and we'll generate a tailored heartbeat checklist.

Your config is ready.

Sign in to unlock your files, setup guide, and Pro extension access.

$49 one-time · Includes 14 days Discord support
← Back to edit
Your config is ready
Show full setup walkthrough — complete terminal commands from install to first message

🔄 How to switch models mid-chat

▼

⚙️ What each config setting does

▼
model.primary
The model that handles most requests (70-80%). Pick the best one your budget allows.
model.fallbacks
Used when primary fails or is rate-limited. Best practice: use a different provider.
heartbeat.every
How often the agent checks for pending work. Default 30m. Uses the cheapest model.
heartbeat.model
The model for heartbeat checks. This fires ~48 times/day — using a premium model here is the #1 cost mistake.
contextPruning.ttl
How long cached context stays valid. 6h is a good default. Lower = less token re-use.
compaction.memoryFlush
When context exceeds the threshold, the agent distills key info to persistent memory files.
maxConcurrent
Prevents runaway loop costs. Limits how many agent turns can run simultaneously.
gateway.bind: "loopback"
Restricts the gateway to 127.0.0.1. Never use 0.0.0.0 — Shodan scans found 135,000+ exposed instances.

➕ How to add more providers later

▼

Run the auth command for any provider:

openclaw models auth login --provider anthropic
Stores key in system keychain (not config file)
openclaw models auth login --provider openai
Same for OpenAI
openclaw models auth login --provider google
Same for Google Gemini
openclaw models auth login --provider openrouter
One key for all models

Then add the model to your fallback chain in openclaw.json and restart: openclaw gateway restart

🔧 Troubleshooting common issues

▼
"command not found: openclaw"
Add npm's bin to your PATH:
echo 'export PATH="$PATH:$(npm config get prefix)/bin"' >> ~/.bashrc && source ~/.bashrc
For macOS/zsh: replace .bashrc with .zshrc
Agent not responding / heartbeat silent
Check openclaw doctor — often a missing models.json in the agent directory. Also check openclaw gateway logs --follow
Token costs unexpectedly high
Check heartbeat model — if it's your primary (expensive) model, that's ~48 calls/day on the wrong model. Also check cron jobs — they may be using stale model config. Run openclaw config set heartbeat.model "google/gemini-2.0-flash-lite"
Tool calls failing silently
Your model may not support tool calling. Claude Sonnet/Opus and GPT-5.x work well. Avoid DeepSeek Reasoner for tool calls. For Ollama: disable streaming (stream: false in provider config).
WhatsApp session expired
Re-scan QR: openclaw channel login whatsapp. Sessions expire periodically. Consider using a dedicated eSIM.
Config changes not taking effect
OpenClaw stores config in 4 places. Stop gateway first → edit config → restart: openclaw gateway stop && openclaw gateway start