Most AI assistants live in a browser tab. They reset between sessions and can't do anything on your machine without you copy-pasting back and forth.

Peter Steinberger wanted something different. He built an AI agent that runs on your own hardware and remembers you across sessions. You talk to it through Telegram while it works on your desktop.

What started as a weekend hack became the fastest-growing repository in GitHub history: 175,000+ stars in under three months, growing 18x faster than Kubernetes. It was called ClawdBot, then MoltBot (Anthropic's trademark lawyers had opinions), and now OpenClaw (as of Feb 16, 2025, OpenClaw has officially joined hands with OpenAI).

I've been running OpenClaw for weeks now, and I'm sharing what I learned. This guide covers what OpenClaw is, how to get started, what the community is building with it, and how to give your agent access to live web data through Firecrawl.

• OpenClaw (formerly ClawdBot, then MoltBot) is a self-hosted AI agent you control from your phone through Telegram, WhatsApp, or other messaging apps
• It runs a local gateway on your hardware and uses LLMs to execute tasks on your machine
• This article covers OpenClaw's origins, what the community is building, personal lessons from daily use, and adding Firecrawl for live web data access

OpenClaw is a self-hosted AI agent you control from your phone through Telegram, WhatsApp, or other messaging apps. It runs on your own hardware with full system access (browser, terminal, filesystem) and remembers you across sessions.

OpenClaw's creator is Peter Steinberger, an Austrian iOS developer. He previously built the company PSPDFKit, a PDF SDK used on over a billion devices, that he sold in 2021 for roughly 100 million euros. He then went through what he described as "profound existential emptiness", and his GitHub activity went dark for three years.

Steinberger came back to coding in April 2025 and started experimenting with AI. By November, he had a working prototype: a WebSocket gateway that relayed WhatsApp messages to Claude's CLI agent. The agent ran shell commands on his machine and sent results back through chat. He named it ClawdBot (a lobster-claw spin on "Claude") and open-sourced it.

The project hit 9,000 stars on its first day. TechCrunch and Hacker News picked it up, and within a week it had more Google searches than Claude Code and Codex combined.

Then Anthropic's trademark team sent a letter ("Clawd" sounded too much like "Claude".) Steinberger renamed it to MoltBot on January 27, 2026, a reference to molting lobsters do to grow. Three days later, he renamed it again to OpenClaw because MoltBot "never quite rolled off the tongue." Reddit called it "the fastest triple rebrand in open source history."

Now, OpenAI has joined hands with OpenClaw.

OpenClaw has two components: a gateway server and an agent runtime. The gateway runs on your machine (Mac Mini, Linux box, even WSL2 on Windows) and connects to Telegram, WhatsApp, Slack, Discord, Signal, iMessage, and a handful of others.

When you send a message, the gateway passes it to the agent, which calls an LLM and executes whatever the model decides to do. It supports Claude, GPT, DeepSeek, Gemini, and local models through Ollama.

OpenClaw runs on your hardware, not in a browser tab. It has full system access: browser automation, email, terminal, filesystem. It also remembers you across sessions through workspace files:

• SOUL.md defines its personality
• USER.md stores your preferences
• AGENTS.md holds its instructions

You talk to it from your phone while it works on another machine.

OpenClaw runs with full access to your machine's terminal, filesystem, and browser. Running it on your primary computer carries real risk. The easiest safe setup for most users is a cloud VM or another cheap, isolated machine you can wipe if anything goes wrong. A DigitalOcean Droplet ($6/month) or Google Cloud e2-micro (free tier eligible) both work well, although the current recommendation is a Mac Mini. I ended up running OpenClaw on EC2. I'll go into more about this later on.

Setup takes about ten minutes. On your chosen VM, install OpenClaw and run the onboarding wizard:

npm install -g openclaw@latest
openclaw onboard --install-daemon

Alternatively, the docs recommend curl to install:

curl -fsSL https://openclaw.ai/install.sh | bash

The wizard walks you through picking a model provider (Claude via Anthropic API or a Pro/Max subscription, GPT, DeepSeek, or a local model through Ollama) and basic security settings.

Once onboarding completes, verify your gateway is actually running before moving on: openclaw gateway status. If it's not, run openclaw gateway start and check for errors.

If anything looks off, openclaw doctor runs a diagnostic check on your gateway, channels, and model provider configuration. For a full walkthrough with screenshots, check the official docs.

Once onboarding finishes, connect Telegram. Open @BotFather in Telegram, send /newbot, and follow the prompts to get a bot token. Then:

openclaw channels login --channel telegram

Paste your bot token when prompted. The gateway connects to Telegram and your bot goes online. Send it a message to verify everything works. Something like "What time is it?" is enough. If the agent responds in chat, you're set.

If any issues arise, use the openclaw doctor.

The two biggest Hacker News threads about OpenClaw (Ask HN: Any real OpenClaw users? and OpenClaw is changing my life) are full of real use cases:

• One user runs six agents as a personal staff: a PA that supervises the others, one handling Twitter growth, another scouting job listings, and a fourth paper-trading crypto with a claimed 77% win rate
• A designer built mold-making software (plastic to silicone to plaster to clay) in about 12 hours
• Someone migrated a Python 2.x home automation setup to Python 3 in 20 minutes
• Others are managing 15,000+ email inboxes, building location-based games you play by walking through parks, and generating Lilypond music arrangements

The consensus among experienced engineers in the threads is that OpenClaw works best for greenfield projects and boilerplate. It struggles with large existing codebases and complex refactoring. Several commenters warned about token costs ($200-400/month for heavy use) and flagged suspicious testimonial patterns in the threads themselves.

The skills ecosystem sits at 5,705+ community-built skills on ClawHub, though a curated list (awesome-openclaw-skills) filters that down to 3,002 after removing 1,180 spam accounts, 672 crypto schemes, 492 duplicates, and 396 flagged as malicious. The 27 remaining categories range from DevOps and smart home control to an AI dating assistant called flirtingbots.

I connected OpenClaw to Telegram and immediately noticed something I didn't expect. Telegram bots support slash commands natively, so OpenClaw surfaces more than 20 of them by default: /new, /reset, /status, /context, /think, and more. It feels closer to using the terminal interface (the most powerful way to run OpenClaw) than a typical chat.

Once I got comfortable with the basics, I started turning it into a personal assistant. I gave it access to Gmail, Google Calendar, and Drive through the built-in skills, so I can manage client work and communication from my phone.

The Telegram setup has a side benefit that it's always there. If I'm texting friends and someone mentions a birthday or a dinner, I switch to my OpenClaw chat, tell it to create a reminder, and go back to the conversation. No opening a separate app.

I also set up heartbeats, which are cron jobs that run on whatever schedule you define. Mine fires every morning with a summary of who's waiting for an email response and what's on my calendar for the day.

But the real effect OpenClaw has had on my life is as an accountability partner. I have ADHD and struggle with executive dysfunction, which means severe time blindness and a hard time meeting deadlines. So I built a system for it.

Using Claude Code (not OpenClaw; my $200 Max account covers coding), I built a local Mac agent called Narc. It captures periodic screenshots of all my displays along with window titles and browser history, then pushes everything to a shared S3 bucket.

OpenClaw polls that bucket on a heartbeat schedule. When it finds me doing something I shouldn't be during an enforced work session (which it can see on my calendar), it confronts me on Telegram. I'm currently working on hooking it up to my actual Telegram account so it can message friends and family when I'm caught slacking. Consequences with witnesses.

There's also a library of community-built skills you can install. Here are a few from the curated list:

• pndr - all-in-one productivity: idea capture, task management, journaling, habit tracking
• cognitive-memory - multi-store memory that mimics human short-term and long-term patterns
• moltbot-ha - controls Home Assistant devices through natural language
• voice-reply - local text-to-speech using Piper voices, no cloud API needed
• quests - guides you through complex multi-step processes like buying a house or filing taxes

But one thing that needs more than a skill install is web search. OpenClaw's web_search doesn't work at all without a provider. The options are a Brave Search API key or Firecrawl. I went with Firecrawl because Brave gives you links, and when the agent tries to actually read those pages, it runs into 403 errors and bot protection on most modern sites. Firecrawl returns the actual page content with search results, so the agent skips that problem entirely.

The simplest way to add Firecrawl is dropping your API key into the config file. This gives your agent a fallback for JS-heavy and bot-protected pages. For web search, crawling, and structured extraction, you install the full Firecrawl skill through Telegram in about 30 seconds.

Open ~/.openclaw/openclaw.json and add your Firecrawl API key:

{
  "tools": {
    "web": {
      "fetch": {
        "firecrawl": {
          "apiKey": "fc-YOUR-API-KEY"
        }
      }
    }
  }
}

With this config, your agent falls back to Firecrawl whenever a page can't be read with a standard HTTP request. JavaScript-rendered sites, anti-bot protection, paywalls: they all go through Firecrawl instead of returning garbage or a 403. You can tune maxAgeMs to control how long cached results stay fresh (default is 2 days).

This covers individual page reads. For web search, site crawling, and structured extraction, you want the full skill.

You don't need to touch the terminal. Just text your agent:

Claw finds the available Firecrawl skills on ClawHub, asks which one you want, and installs it. The firecrawl-skills option gives your agent scrape, search, crawl, and map under the hood. It authenticates automatically if you have FIRECRAWL_API_KEY set in your environment.

From here, you just talk to your agent like normal. When it needs to read a webpage, it scrapes through a real browser instead of failing on JavaScript. When it needs to research something, it searches the web and pulls back full page content instead of just links. When you point it at a documentation site, it can crawl every page and map the structure.

Here's what that looks like in practice. I asked my agent to compare pricing across three SaaS products:

It searched the web, scraped each pricing page, and put the results together in one message. No URLs needed, no 403 errors, no browser tabs.

You can learn more at our Guide to the Firecrawl Skill and CLI.

I started running OpenClaw on my MacBook because I don't have a desktop. That lasted about a week. Reports about OpenClaw's security vulnerabilities started piling up (exposed control panels, one-click RCE, credentials stored in plaintext), and running an agent with full system access on my primary machine didn't sit right.

On top of that, keeping a laptop open 24/7 isn't practical. Battery health, sleep settings, lid behavior. I uninstalled completely.

The fix was deploying to AWS. I spun up an EC2 instance (t4g.small, about $15/month) and moved everything there. The agent was now up around the clock, and my accountability system with Narc could poll S3 without me worrying about my MacBook sleeping mid-session.

If you're considering a Mac Mini for this (the common recommendation), I'd argue EC2 is a better deal unless you need the macOS-specific features. A t4g.small with an existing Anthropic subscription runs under $30/month total.

Setting up the EC2 was the hardest part because you install everything from scratch: Node.js, OpenClaw, systemd services, security groups. Here's a single script that handles the full deployment. It assumes your AWS credentials are configured locally with the right permissions:

#!/bin/bash
# deploy-openclaw.sh
set -euo pipefail
 
REGION="us-east-1"
INSTANCE_TYPE="t4g.small"  # 2 vCPU, 2GB RAM, ~$15/month
KEY_NAME="openclaw-key"
 
# SSH key pair
aws ec2 create-key-pair \
  --key-name "$KEY_NAME" \
  --key-type ed25519 \
  --region "$REGION" \
  --query 'KeyMaterial' --output text > "${KEY_NAME}.pem"
chmod 400 "${KEY_NAME}.pem"
 
# Security group: SSH only, gateway port stays closed
SG_ID=$(aws ec2 create-security-group \
  --group-name openclaw-sg \
  --description "OpenClaw server" \
  --region "$REGION" \
  --query 'GroupId' --output text)
 
MY_IP=$(curl -s https://checkip.amazonaws.com)
aws ec2 authorize-security-group-ingress \
  --group-id "$SG_ID" \
  --protocol tcp --port 22 \
  --cidr "${MY_IP}/32" \
  --region "$REGION"
 
# Ubuntu 24.04 ARM64
AMI_ID=$(aws ssm get-parameter \
  --name /aws/service/canonical/ubuntu/server/noble/stable/current/arm64/hvm/ebs-gp3/ami-id \
  --query 'Parameter.Value' --output text \
  --region "$REGION")
 
# Launch
INSTANCE_ID=$(aws ec2 run-instances \
  --image-id "$AMI_ID" \
  --instance-type "$INSTANCE_TYPE" \
  --key-name "$KEY_NAME" \
  --security-group-ids "$SG_ID" \
  --block-device-mappings 'DeviceName=/dev/sda1,Ebs={VolumeSize=30,VolumeType=gp3}' \
  --tag-specifications 'ResourceType=instance,Tags=[{Key=Name,Value=openclaw-server}]' \
  --region "$REGION" \
  --query 'Instances[0].InstanceId' --output text)
 
aws ec2 wait instance-running --instance-ids "$INSTANCE_ID" --region "$REGION"
 
EC2_IP=$(aws ec2 describe-instances \
  --instance-ids "$INSTANCE_ID" \
  --region "$REGION" \
  --query 'Reservations[0].Instances[0].PublicIpAddress' --output text)
 
echo "Instance up at ${EC2_IP}. Waiting for SSH..."
sleep 30
 
# Install Node.js, OpenClaw, and start the gateway
ssh -o StrictHostKeyChecking=no -i "${KEY_NAME}.pem" ubuntu@"${EC2_IP}" << 'REMOTE'
set -euo pipefail
 
sudo apt-get update && sudo apt-get upgrade -y
 
# Node.js 22 via NVM
curl -o- https://raw.githubusercontent.com/nvm-sh/nvm/v0.40.1/install.sh | bash
export NVM_DIR="$HOME/.nvm"
[ -s "$NVM_DIR/nvm.sh" ] && . "$NVM_DIR/nvm.sh"
nvm install 22
 
npm install -g openclaw@latest
 
# Headless systemd support (required on EC2)
sudo loginctl enable-linger ubuntu
export XDG_RUNTIME_DIR=/run/user/$(id -u)
echo 'export XDG_RUNTIME_DIR=/run/user/$(id -u)' >> ~/.bashrc
 
# Onboard and start
openclaw onboard --non-interactive --gateway-bind loopback --skip-daemon --skip-skills
openclaw gateway install --force
openclaw gateway start
 
echo "OpenClaw running. Set your API keys and connect Telegram."
REMOTE
 
echo "SSH: ssh -i ${KEY_NAME}.pem ubuntu@${EC2_IP}"
echo "Tunnel: ssh -L 18789:127.0.0.1:18789 -i ${KEY_NAME}.pem ubuntu@${EC2_IP}"

Once the instance is up, SSH in, set your ANTHROPIC_API_KEY (or integration token) and FIRECRAWL_API_KEY in ~/.bashrc, then connect Telegram with openclaw channels login --channel telegram. The gateway port (18789) stays closed to the public. You access it through an SSH tunnel if you ever need the web interface.

After a few weeks of running, I hit a problem. Claw's responses started getting slower and slower. The typing indicator would hang for minutes, even on short messages. Starting a fresh conversation with /new didn't help. I ran openclaw doctor --deep and did a full audit of the system.

The cause: without regular cleanup, heartbeat cron jobs and conversation history accumulate stale data. OpenClaw was injecting all of it into every new thread, burning through my Sonnet 4.5 account limits on context that didn't matter anymore.

I was very glad to have used a Claude integration token instead of a raw API key. With per-token billing, that kind of silent context bloat would have been an expensive surprise.

OpenClaw went from a one-hour WhatsApp relay to 175,000+ GitHub stars in under three months. Peter Steinberger joined OpenAI on February 15, 2026, and the project is moving to an open-source foundation with OpenAI's continued support.

For me, it started as a curiosity and turned into something I use every day. The accountability system, morning email digests, calendar reminders mid-conversation, Firecrawl-powered web research... all running on a $15/month EC2 instance that I set up in an afternoon. The rough edges are real (stale data buildup, security concerns, the triple rebrand), but nothing that made me want to stop using it.