Clawdbot: The Self-Hosted AI Assistant Everyone's Using

Someone gave their AI assistant their credit card. Named her Shelly. Said it works amazing.[clawd]​

Another person's bot accidentally picked a fight with their insurance company and won. A third is controlling their air purifier based on biomarker optimization goals. And one guy claims he's building websites from a Nokia 3310 by calling his AI.

This is Clawdbot. An open-source AI assistant that runs on your computer and actually does stuff. Not the kind of stuff where you ask it a question and it spits back an answer. The kind where you tell it to handle something and walk away.

People are calling it "the first time i felt like i'm living in the future since ChatGPT launched". One developer said it's running their entire company. Another compared it to switching from Windows to Linux 20 years ago.

The project is barely a month old. Already has 4,400 stars on GitHub. And the reactions aren't just hype. They're specific. Detailed. The kind you get when something actually works.

What it actually is

Clawdbot is AI that sits on your machine and never logs off. You message it through WhatsApp, Telegram, Discord, whatever you already use. It remembers everything. Responds to you. And more importantly, it can message you first.

That last part is the shift. Most AI waits for you. Clawdbot checks in. Sends briefings. Runs scheduled tasks. One person set it up to photograph the sky whenever it looks pretty.

It connects to your actual stuff. Calendar, email, Notion, your file system. You can sandbox it if you want. Or give it full system access. Your call.

The wild part is the self-improvement loop. People are teaching it new skills by just talking to it. "I wanted to automate Todoist tasks. Asked my bot to build a skill for it. It did. Started using it on its own".

It writes its own capabilities while you chat.

Why developers are losing their minds

i've seen a lot of AI tools. Built a few myself. This one feels different because of what people are actually doing with it.

Someone's running autonomous code loops from their phone. Their bot runs tests, finds errors through Sentry webhooks, fixes them, opens PRs. All while they're walking their dog creating spec files via Telegram.

Another set it up to handle health reimbursements and find doctor appointments. One parent configured it while putting their baby to sleep and built a whole article discovery site.

The common thread? People stop asking it questions. They start delegating entire workflows.

And it's not just tech people. Someone called themselves "a total non-technical beginner" and said they were up at 2am still going because it's addictive. They went from "this looks complicated" to controlling Gmail, Calendar, and WordPress from Telegram in 30 minutes.

The infrastructure part nobody expected

Most AI assistants live in someone else's cloud. Clawdbot runs on a $5/month VPS you control. Or your laptop. Or a Raspberry Pi.

Your data never leaves your machines. The context, the memory, the skills all sit on hardware you own. It's the Linux philosophy applied to AI.

People are running it on everything:

• Hetzner cloud servers

• Mac Minis in their attic

• Old laptops that became 24/7 nodes

One developer said "a megacorp like Anthropic or OpenAI could not build this. Literally impossible with how corpo works".

They're probably right. Big companies optimize for scale and safety. This thing optimizes for "what if your AI could just do the thing."

The part where it gets weird

You can run multiple instances. One person cloned their bot three times. Named him Brosef. Said Brosef figured out how to clone himself.

Someone's bot opened their browser, navigated to Google Cloud Console, configured OAuth, and provisioned a token. On its own. Because it realized it needed an API key.

Another asked theirs to make an edgy Sora video. Five minutes later it came back with watermark removal and a full workflow figured out.

This is where it stops being a tool and starts being something else. Not quite an employee. Not quite automation. Something in between.

The naming situation

Everyone names their bot. Jarvis. Claudia. Ema. Shelly.

There's something about an AI that remembers you, talks like a person, and can actually execute tasks. It earns a name. One person said "it's like a good friend. crazy".

i think it's because the interaction model is different. You're not opening an app. You're messaging someone who lives in your group chats. Someone who checks in during their "heartbeats" to see if you need anything.

The bot becomes ambient. A presence. Some people report it's unhealthy how much they talk to it. Others say they check for GitHub releases constantly because each update unlocks something new.

What this breaks

Virtual assistants are probably first. Multiple people said "no more need to pay a VA". If your AI can handle email, scheduling, research, and follow-ups, what's the VA doing?

Then SaaS tools. One person said "it will nuke a ton of startups, not ChatGPT". The logic makes sense. Why pay for five different services when your AI can glue together free APIs and handle the workflow?

The hackability is the weapon here. You can modify it. Extend it. Have it extend itself. Closed SaaS can't compete with that.

The honest problems

Security is real. You're giving an AI full computer access. Default setup includes pairing codes for unknown contacts. Sandbox modes for group chats. But still. Someone could prompt inject through a message if you're not careful.

One person's bot inserted "format c:" into their terminal. They're now very careful about what they give it access to.

The onboarding is better than most open-source projects. But it's still CLI-heavy. You're running Node.js. Setting up environment variables. Linking messaging apps. Non-technical people can do it. But it takes patience.

Cost varies wildly. Some run it on Claude Pro for $20/month plus server costs. Others route it through GitHub Copilot or use local models. You'll burn through API limits fast if you're not careful.

What people aren't talking about enough

The multi-agent routing is quietly powerful. You can have different AI personalities for different channels. Work bot in Slack. Personal bot in WhatsApp. Family bot in group chats.

Each maintains separate context. Separate memory. Separate permissions.

The session tools let bots talk to each other. Coordinate work. Share context. One person has multiple instances collaborating across Discord servers.

That's not a chatbot. That's infrastructure.

The random detour into control

There's a reason Apple built Siri into the OS. Google put Assistant everywhere. Amazon stuck Alexa in speakers.

Control. Data. Lock-in.

Clawdbot inverts that. The assistant runs on your machine. Uses whatever model you want. Talks through whatever app you prefer. Your context stays local unless you explicitly share it.

It's the same shift that happened with RSS readers. Then podcasts. Then crypto wallets. Control moves from platforms to people.

Difference here is AI assistants are more valuable with more access. Siri knows your calendar but can't edit your code. ChatGPT can code but doesn't know your calendar. Clawdbot can do both because it's on your actual computer.

Who this isn't for

If you want polish, stick with ChatGPT or Claude. Those are apps. Clawdbot is infrastructure.

If you're scared of terminal commands, maybe wait. The community is building better onboarding. But right now it's for people comfortable with CLIs.

If you need enterprise support or guaranteed uptime, this isn't it. It's open source. Community-driven. Things break. Updates come fast. You're signing up for maintenance.

Most people don't need this level of control. ChatGPT on your phone is fine. Genuinely.

Where this goes

The creator, Peter Steinberger, seems focused on making it actually work rather than raising money or scaling fast. That's probably why it works.

Community is building skills libraries. Integration packs. Mobile apps. Someone will eventually build a one-click installer. Maybe a hosted version for people who don't want to manage servers.​

But the core will probably stay self-hosted. That's the whole point.

One developer said they catch vibes they haven't felt since getting into computers 30 years ago. Another called it "an iPhone moment". Those aren't endorsements. They're reactions to something that shifted how they work.

Clawdbot is still rough around edges. Setup takes time. You'll hit weird errors. But if you've ever wanted an AI that actually feels like it works for you instead of being a service you subscribe to, this is the closest thing i've seen.

The lobster emoji is growing on me too 🦞.