ClawdBot AI: Installation Guide, Use Cases & Expert Tips

ClawdBot AI started as a weekend experiment by iOS developer Peter Steinberger in January 2026 and exploded within weeks into one of the most talked-about open-source projects of the year. The concept was simple: what if your AI assistant lived inside WhatsApp or Telegram, remembered everything you told it, and could actually do things — send emails, control your browser, run shell commands — instead of just answering questions? Over the course of four weeks, the project was renamed twice (Clawdbot, then Moltbot, then OpenClaw) and spawned a whole ecosystem of alternatives. The core idea stuck: a self-hosted AI agent that runs on your hardware, connects to the messaging apps you already use, and works for you 24/7. This guide walks through installation, configuration, real use cases people are running, and the quirks you’ll hit along the way. Last verified from official docs April 2026.

⚠️ Heads-up: ClawdBot gives an AI agent real access to your computer — filesystem, shell, browser. That’s powerful and risky. Start with the permission system locked tight, test on a non-personal machine or fresh user account, and read the security section before running it on anything sensitive.

📋 What this guide covers: The full install flow for macOS, Linux, and Windows (via WSL2), how to connect your first messaging channel, the onboarding wizard, the actual use cases people are running day-to-day, and the troubleshooting problems that come up most. All steps verified against the official documentation and GitHub repo. The project was named ClawdBot at launch, then Moltbot, now officially OpenClaw — we use ClawdBot throughout this guide since that’s what most people still search for.

⚡ TL;DR — ClawdBot AI at a Glance

• 🧠 What it is: Open-source self-hosted AI agent that connects to WhatsApp, Telegram, Discord, Slack, Signal, iMessage
• ⚙️ How it works: Runs a Gateway daemon on your machine, routes messages between channels and an LLM (Claude, GPT-4o, or local via Ollama)
• 💰 What it costs: Software is free. You pay only for the LLM API — roughly $0.10 to $1 per conversation turn with Claude, $0 with local Ollama models
• ⏱️ Setup time: 20 to 40 minutes for first-time installs
• 🖥️ Where to run it: Best on an always-on machine — Mac Mini, Raspberry Pi, VPS, or spare laptop
• 🔗 Official links: docs.openclaw.ai and the openclaw GitHub organization

Table of Contents

Prerequisites — What You Need Before You Start

Step 1 — Install Node.js

ClawdBot needs Node.js version 22 or newer. Check what you’ve got by running node –version in your terminal. If it outputs v22 or higher, skip ahead. If not, the recommended install path is through nvm (Node Version Manager) — the install script is on the nvm GitHub page, and once it’s installed you run nvm install 22 followed by nvm use 22.

On macOS, Homebrew works fine too: brew install node@22. On Windows, install WSL2 first (open PowerShell as admin and run wsl –install -d Ubuntu), restart, then run the Linux instructions inside your Ubuntu shell.

Step 2 — Install ClawdBot CLI

The one-line installer (available at clawd.bot/install.sh) is the simplest route. It detects your OS, installs dependencies, and kicks off the onboarding wizard. If you’d rather install via npm directly, run npm install -g clawdbot@latest, then confirm the install by running clawdbot –version — it should print a version string.

If you hit build errors on native dependencies (sharp is the usual suspect), set the environment variable SHARP_IGNORE_GLOBAL_LIBVIPS=1 before re-running the install. For cloud servers or Raspberry Pi setups, the official Docker image at ghcr.io/steipete/clawbot:latest is the path of least resistance — mount a volume for the config directory and expose port 18789.

Step 3 — Run the Onboarding Wizard

With the CLI installed, start the interactive onboarding flow by running clawdbot onboard –install-daemon. This is where the bot asks what LLM you want, which messaging channels to connect, and whether to install itself as a background daemon. The –install-daemon flag is important — without it, the bot only runs while your terminal is open.

The wizard walks through five steps:

1. Gateway mode: Pick “local” for single-device setups, “remote” if running on a VPS or cloud box.
2. Authentication: Set an API key for whatever LLM backend you’re using.
3. Model selection: Claude Opus is the top pick for complex agent work. Local Ollama models (Llama 3, Mistral, DeepSeek) work offline but with reduced reasoning quality.
4. Channel setup: Telegram is the easiest first channel — create a bot via @BotFather in Telegram, paste the token when prompted. WhatsApp requires a QR scan from your phone.
5. Daemon install: Confirm persistent background running.

Test the connection by messaging your bot in Telegram: “Hello, what’s my status?” You should see a reply within a second or two.

💡 Expert Tip — Use a dedicated Apple ID or OS user: ClawdBot has filesystem and shell access on the machine where it runs. Don’t give it access to your personal iCloud, browser profiles, or saved passwords. On a Mac Mini, create a fresh Apple ID and system account dedicated to the bot. On Linux, a dedicated user account with restricted sudo rights does the same job. This is one of those “obvious in hindsight” steps that matters more than the dramatic security tips.

Step 4 — Gateway and Channels

After onboarding, the Gateway (the core service) runs on port 18789 by default. The useful commands you’ll run regularly: clawdbot gateway status to check whether it’s running, clawdbot gateway restart or clawdbot gateway stop to control it, and clawdbot logs –follow to tail the logs in real time. To see your connected messaging channels, clawdbot channels status lists them with their health state.

To add a second channel later (say WhatsApp after starting with Telegram), run clawdbot channels login and scan the QR code with WhatsApp Business on your phone. You’ll find the pairing option under Settings → Linked Devices → Link a Device.

Real-World Use Cases People Are Running

ClawdBot works best when you treat it like a staff member with read/write access to your digital life, rather than a chatbot. Here’s what the community actually builds with it:

Daily life automation

• 📧 Inbox triage: “Clear my inbox of newsletters” — parses Gmail labels, files or archives based on rules you describe in plain English.
• 📅 Calendar management: “Schedule a meeting with Priya next Tuesday at 3 PM if I’m free” — checks your calendar, creates the invite, confirms by Telegram message.
• 📊 Morning briefing: “Every morning at 8 AM, summarize my calendar and top five emails” — cron job runs daily, arrives in Telegram before you reach for coffee.

Browser and web automation

• ✈️ Flight hunting: “Find the cheapest direct flight to Berlin next month, report back with 3 options” — bot launches browser, searches Google Flights, summarizes.
• 🛒 Order tracking: “Check the status of my Amazon orders from last week” — logs into your saved session, pulls the list, returns it.
• 📰 Research summaries: “Read the top 10 Hacker News posts and summarize the most interesting three” — common morning routine for technical users.

File and shell work

• 📁 Quick summaries: “Summarize my PDF report.pdf” — reads the file from disk, returns key points.
• 🧹 Folder organization: “Organize my downloads folder by file type” — creates folders, moves files, reports what it did.
• 🔧 Dev ops shortcuts: “Run git pull on my project repo and show me the changed files” — bot asks for permission on the first run, then executes and reports.

Crypto and markets

• 📈 Price alerts: “Alert me if BTC moves more than 3% in an hour” — bot watches an API, pings Telegram when triggered.
• 🪙 On-chain monitoring: “Monitor wallet 0xabc for any new outgoing transactions” — uses a custom skill, notifies on events.
• 📰 News summaries: “Summarize what happened in crypto overnight while I was asleep” — pulls from configured news sources, delivers a morning digest.

💡 Expert Tip — Write custom skills for anything repetitive: The biggest unlock with ClawdBot isn’t the bundled features — it’s custom skills. Ask the bot “Create a skill to monitor Solana memecoin launches and alert me when a new token hits 1,000 holders in its first hour.” It writes the code, you approve it, and from then on that skill is available as a native command. Most power users end up with 20 to 40 custom skills covering their specific workflows within the first month.

Model Selection — Claude vs GPT vs Local

ClawdBot is backend-agnostic. The three common choices behave differently for agent work:

Most users start with Claude Opus because tool-use reliability matters more than anything else for agent work. Local models are worth setting up as a fallback for privacy-sensitive tasks.

Expert Tips & Tricks

• 🧠 Use /think high for complex tasks. This triggers deeper reasoning mode — slower and more expensive but much better for multi-step plans.
• 🎙️ Enable ElevenLabs TTS for audio replies. Useful on mobile when you’re walking or driving. Set it per-channel so your desktop stays text-only.
• 🔐 Run dangerous commands in consent mode first. Until you trust a skill, keep confirm-shell mode on. The bot will ask before running destructive commands.
• 🧑‍🤝‍🧑 Separate work and personal agents. Multi-agent routing lets you set up one bot for work (with access to your work Gmail, Slack) and another for personal life. Don’t mix them.
• ⏰ Lean on cron. Most of the value comes from scheduled jobs — morning briefings, weekly summaries, market alerts. Set them up early and add as you go.
• 🦺 Keep the permission system locked for the first month. Every new skill should require explicit approval the first time it runs. You can loosen this later once you trust specific skills.
• 💾 Back up your config folder. It contains your agent’s memory, skills, and channel keys. Losing it means starting over.

Common Problems & Troubleshooting

Gateway won’t start on port 18789: Another service is using the port. Restart with a different port using clawdbot gateway –port 19000. Also check if a previous instance is still running.

Telegram messages never arrive: Your bot token is wrong or the bot was disabled in @BotFather. Regenerate the token, update your config, restart the gateway.

WhatsApp keeps disconnecting: WhatsApp Business kicks idle sessions. If you’re running on a laptop, disable sleep. On a Mac Mini, go to System Settings → Lock Screen → never turn off display while plugged in.

“sharp” errors during npm install: Set the SHARP_IGNORE_GLOBAL_LIBVIPS environment variable and retry. If that fails, you’re probably missing build tools — install node-gyp.

Bot replies slowly (5+ seconds): Usually LLM latency, not ClawdBot. Switch from Claude Opus to Haiku for routine queries, or from a remote API to local Ollama if you care more about speed than quality.

High API bills: Set a daily budget cap in your config. Most per-conversation runaway costs come from loops — the bot calling itself or running cron jobs more often than intended. Check the logs frequently.

Security — What to Lock Down Before Going Live

Because ClawdBot can read files, run shell commands, and browse the web on your behalf, the attack surface is real. Treat it like giving an intern unsupervised access to your machine:

• 🛡️ Dedicated user account. Never run it as your personal user or as root.
• 🔑 Isolate API keys. Put the Anthropic API key in a dedicated account with a billing cap, not your main one.
• 📵 Restrict shell access by default. Turn on confirm-shell mode in the config. You can always approve a command on the fly, but accidental file deletion is forever.
• 🚪 Two-factor your messaging accounts. If someone gets into your Telegram, they can message your bot as you. 2FA on Telegram and WhatsApp is non-negotiable.
• 🧾 Review logs weekly. The log output shows every command the bot executed. If anything looks off, find out why.

Who ClawdBot Is (and Isn’t) For

• ✅ Great fit: Developers, technical users, traders, anyone comfortable with Node.js and a terminal. If cloning a repo and running pnpm install doesn’t intimidate you, you’ll be fine.
• ✅ Great fit: Privacy-minded users who want local control over an AI agent that knows their schedule, emails, files.
• ❌ Not ready for: Non-technical users. The onboarding wizard has come a long way, but you’ll hit a terminal at some point and need to solve problems there.
• ❌ Not ready for: Production workloads in regulated industries. This is still a project in rapid evolution — breaking changes between versions are common.

Bottom Line: ClawdBot is the most compelling self-hosted AI agent project running today — and the fact that the whole category basically didn’t exist four months ago tells you how early we are. The value comes from the persistent memory, the messaging-channel integration, and the custom-skill system. Install it on a dedicated always-on machine, start with Telegram and Claude Opus, keep permissions locked for the first month, and build your custom skill library from the repetitive work you’re doing manually right now. Expect to spend a weekend on the initial setup and tuning. After that, it fades into the background and quietly makes you faster. For ecosystem alternatives that are lighter, more security-focused, or hardware-optimized, see our comparison of Nanobot, NanoClaw, IronClaw, ZeroClaw, PicoClaw and TinyClaw.

FAQs

Is ClawdBot free? The software itself is open-source and free. Your costs are the LLM API (roughly $3 to $10/day with Claude Opus on heavy use, $0 with local Ollama models) and your hardware. Most people who run it say they spend $30 to $100/month total.

Can I run ClawdBot on Windows? Yes, through WSL2. Native Windows isn’t officially supported yet. Install WSL2 in PowerShell as admin, restart, then follow the Linux install steps inside your Ubuntu shell. Keep your PC awake while the bot is running.

What’s the difference between ClawdBot, Moltbot and OpenClaw? Same project — it got renamed twice. Clawdbot (Jan 2026) → Moltbot (late Jan) → OpenClaw (Feb 2026 onwards). The official name now is OpenClaw. Most tutorials and community discussions still use Clawdbot because that’s the name it went viral under.

Does ClawdBot work offline? The ClawdBot software runs locally, but LLM reasoning is not offline unless you run a local model through Ollama (Llama 3, Mistral, DeepSeek). Configure Ollama in the wizard and you can run the full stack without internet access.

Is it safe to give the bot shell access? It’s as safe as you make it. Run it in a dedicated user account with restricted privileges, keep confirm-shell mode on, and review the logs regularly. Don’t run it as root or your main personal account.

Can multiple people share one ClawdBot instance? Yes — connect multiple Telegram users or a WhatsApp group as channels. Permissions and access control are per-channel, so you can give your partner read-only access to the household calendar while keeping shell access restricted to yourself.

What hardware do I need? Base recommendation is an always-on machine with 4 GB RAM minimum, 8 GB comfortable. A Mac Mini M-series is the sweet spot ($500 to $600). Raspberry Pi 5 works for lighter use. A cheap VPS ($5 to $10/month) is fine if you don’t need to connect iMessage (which requires macOS).

How is ClawdBot different from ChatGPT? ChatGPT is a chatbot — it answers questions. ClawdBot is an agent — it takes action in your environment (files, shell, browser, messages). It also has persistent memory across conversations and runs on your hardware, so conversations never leave your network when using local models.

📋 Related: PolyCop Telegram Bot Review | Nanobot vs NanoClaw vs IronClaw vs ZeroClaw vs PicoClaw vs TinyClaw | Best AI Tools for Startups | Best AI Coding Tools