From beginner to advanced, cloud to local — everything you need to raise your OpenClaw
OpenClaw is an open-source AI Agent framework that lets you run a team of AI workers on your own machine. Command them remotely via Telegram, Discord, LINE, and more. They can:
There's more than one AI Agent solution out there. Which one is right for you?
Use Claude Code CLI directly — ask and receive. Great for short tasks, but no persistent memory, no scheduling, and it's gone when you close it.
The platform handles everything — convenient but limited. Low customizability, your data lives on someone else's cloud, and features are capped by the platform.
Fully autonomous, open-source, and free. Your data never leaves your machine. Run multiple Agents simultaneously, each with their own role, running 24/7. Steeper learning curve, but the highest ceiling.
Different models have different strengths. Pick the right one and work smarter.
| Claude Sonnet | Claude Opus | GPT-4o | Gemini 2.5 Pro | |
|---|---|---|---|---|
| Role | Daily workhorse | Highest quality | General purpose | Long context |
| Speed | Fast | Slower | Fast | Fast |
| Cost | Medium | High | Medium | Medium |
| Best for | Everyday tasks | Complex reasoning / writing | Multimodal | Large datasets |
Setting up OpenClaw has two main cost areas: hardware and API.
| Device | Price | Suitable For |
|---|---|---|
| Mac mini M4 16GB | ~NT$20,000 | Entry-level, run 2–3 Agents |
| Mac mini M4 Pro 24GB | ~NT$35,000 | Advanced, multi-Agent + browser automation |
| VPS (Zeabur / Railway) | ~NT$0–500/mo | Dipping your toes in |
| Plan | Monthly | Notes |
|---|---|---|
| Claude Pro | $20/mo | Good for light personal use |
| Claude Max 5x | $100/mo | For power users |
| API Pay-as-you-go | Usage-based | Most flexible, ideal for multi-Agent setups |
Multi-Agent + model tiering can save 50%+ on token usage. Use cheap models for simple tasks, expensive ones for the important stuff.
Step one: decide where OpenClaw runs.
Start with cloud to test the waters. Once you're committed, grab a Mac mini.
From zero to one — just follow along.
Mac / Linux / Windows (WSL2) / VPS all work. On Mac, press Cmd+Space and type "Terminal" to open it. Windows requires WSL2; native Windows is not supported.
Paste this in your terminal: curl -fsSL https://fnm.vercel.app/install | bash && fnm install --lts
npm install -g openclaw@latest
openclaw onboard --install-daemon
The onboarding flow lets you select a provider (Anthropic / OpenAI / Google / xAI / Bedrock, etc.) and installs the background daemon.
openclaw start — when you see 🦞, you're good to go!
Open Telegram / Discord and say hello to your AI agent!
Understand these concepts and you'll understand OpenClaw's architecture.
Your AI worker. Each Agent has its own name, personality, tools, and memory. You can have one for development, one for marketing, one for research.
An instruction manual for your Agent. Install a Skill, gain an ability. The community has hundreds of ready-made Skills to install.
Your Agent's long-term memory. It remembers your preferences, lessons learned, and key decisions. The more you use it, the better it knows you.
A scheduled self-check mechanism that runs automatically. Like an alarm clock that periodically triggers your Agent to inspect, report, and maintain.
The USB-C of the AI world. Lets your Agent connect to external tools like Notion, GitHub, Figma, and more.
Set a time, and your Agent runs automatically. Daily reports at 9 AM, weekly backups on Friday — all automated.
7 releases in the past 10 days (as of 2026/04/09, latest v2026.4.9). Stay updated for security fixes.
New /tasks command. Manage background tasks directly in any chat interface, no IDE switching needed.
Durable flow state, managed/mirrored sync, CLI tools openclaw flows for inspection and recovery.
Built-in video_generate and music_generate tools, integrated with Google Lyria, MiniMax, and ComfyUI workflows.
Control UI now supports 12 languages including English, Traditional/Simplified Chinese, Japanese, Korean, French, German, Spanish.
Added Qwen, Fireworks AI, StepFun, MiniMax, Ollama Web Search, Google Gemma 4, Bedrock Mantle, and more.
New command openclaw infer unifies provider-backed inference (model/media/web/embedding).
memory-wiki plugin with claim/evidence structure, freshness-weighted search, and Dream Diary timeline UI.
External automations can trigger TaskFlow via per-route shared-secret endpoints.
SSRF quarantine, dotenv hardening, dangerous skill fail-closed, node pairing approval. Upgrade ASAP.
May highlights: voice + realtime + multi-platform deepening (as of 2026/05/09, latest v2026.5.9-beta.1).
Discord voice channels support realtime speech-to-text + text-to-speech + agent-proxy. Talk to your lobster directly with voice.
Google Meet calls + Twilio dial-in integrated with Gemini realtime voice bridge, paced audio streaming.
Active Memory plugin supports concrete tool names for custom memory configurations — more precise auto-trigger.
New /think default + /fast default let you switch reasoning depth mid-chat without restarting the session.
iOS app supports setup codes to connect to private LAN gateways — no more tedious manual config.
WhatsApp LID-addressed contacts, Feishu native topic threads, Discord heartbeat ACK timing all hardened.
Unified provider registration for text / image / video / music — new models go live faster.
Plugin publish retry mechanism, runtime fetch header normalization, external plugin install migration hints.
Common commands and config files — bookmark this for quick reference.
/statusCheck system status/modelSwitch models/tasksBackground task board (v2026.4.1+)openclaw startStart OpenClawopenclaw stopStop OpenClawopenclaw restartRestart OpenClawopenclaw flowsInspect/recover Task Flowopenclaw inferProvider inference hub (v2026.4.7+)openclaw doctor --fixAuto-migrate configopenclaw cron --toolsCron tool whitelistMEMORY.mdAgent long-term memory fileHEARTBEAT.mdHeartbeat check configurationAGENTS.mdAgent role definitionsOnce OpenClaw is up and running, unlock more powerful capabilities.
Run multiple agents, each with their own role: one for development, one for marketing, one for research. Independent memory, independent models, no interference.
Let your Agent control the browser: search the web, fill forms, take screenshots, scrape data — like having an assistant sitting right next to you.
Connect Claude Code / Codex and let your Agent write code, open PRs, and run tests automatically. All you do is merge.
Telegram / Discord / LINE / Slack — command your AI team from your phone, anytime, anywhere.
Through Self-Reflection, Memory, and Insight mechanisms, your Agent learns from mistakes and gets stronger over time.
Wisdom from experienced OpenClaw users — skip the common pitfalls.
Don't chase the perfect setup from day one. Install it, start using it, and fine-tune as you go.
Don't make a single Agent do everything. Clear division of labor means higher efficiency.
Your Agent's quality depends on the quality of memory you give it. Invest time in writing solid MEMORY.md and AGENTS.md files.
Don't build everything from scratch. ClawHub has hundreds of ready-made Skills. Stand on the shoulders of giants.
API fees add up. Use model tiering (cheap models for simple tasks, expensive ones for the important stuff) to keep costs in check.
High-quality tutorials and shares from the community, continuously updated.
Pick what fits your needs — we've put together the packages for you.