Note: The tool mentioned in this article, formerly known as Clawdbot, has been rebranded to OpenClaw. This article has been updated to reflect the new name and installation instructions. For the latest information, please visit the official website: openclaw.ai.
Over the past three weeks, I’ve delved into the world of self-hosted AI assistants, specifically OpenClaw (formerly Clawdbot). This article shares my experiences, a guide on how to set it up on a Virtual Private Server (VPS), what models to use, and the most cost-effective way to run it.
My journey with OpenClaw began from a unique need. While traveling, I often found myself on flights with Wi-Fi restricted to messaging services only. I wanted a way to access a powerful AI during these flights. Around that time, OpenClaw was gaining traction, so I decided to experiment. After an initial local setup on my MacBook proved its potential, the logical next step was to deploy it on a device that could run 24/7.
Choosing Your Hosting: VPS vs. Dedicated Hardware
I opted for a VPS, hosting it in the cloud. Many users consider buying a dedicated Mac Mini for OpenClaw. If you have the money to spare, it’s a viable option, but it means another piece of hardware you need to keep powered on and maintained.
So, what’s the real difference between a VPS and a dedicated Apple device?
The primary distinction lies in ecosystem integration. If you’re heavily invested in Apple’s ecosystem, you’ll likely want OpenClaw to access iMessage, iCloud Drive, and Apple Notes. Running it on a non-Apple device, like a Linux VPS, makes this integration incredibly difficult. This has been a pain point for me, but so far, the need hasn’t been strong enough to justify purchasing a separate Mac Mini. For now, I’m content with OpenClaw accessing my non-Apple services like email.
Setting Up on a VPS: A Step-by-Step Guide
Getting OpenClaw running on a VPS is surprisingly straightforward. While many recommend Hetzner, a German hosting company, I found their verification process stringent and heard reports of accounts being banned quickly. Since I already had other services running there, I sought a different provider.
My research led me to Contabo. I’ve been using their service for the 21 days I’ve had OpenClaw running, and it has been flawless and affordable. I’m on their lowest-tier plan, which costs around $4.95 a month, and it’s more than sufficient to run OpenClaw. I even use the same VPS for remote coding sessions from time to time.
Here’s a general workflow for setting it up:
- Get Your VPS: Sign up for a provider like Contabo and purchase a VPS. They will email you the necessary details, including your server’s IP address.
- Establish an SSH Connection: You’ll need to create an SSH connection string to access your server. If you’re unsure how, you can easily ask an AI like ChatGPT for the correct format.
- Use an SSH Client: I use a popular tool called Termius for SSHing into remote devices. The first-time setup is self-explanatory, guiding you through the process of adding a new host.
- Use a Terminal Multiplexer (Optional but Recommended): Once you’re connected to your VPS, I highly recommend using a tool like
tmux. OpenClaw is still early-stage software and can run into issues frequently, requiring you to SSH in and troubleshoot. Withouttmux, each time you connect, you start a fresh terminal session. Withtmux, your session is retained in memory, allowing you to disconnect and reconnect, picking up exactly where you left off. It’s a significant quality-of-life improvement.
Once you’re in the terminal, create a dedicated folder for OpenClaw and follow the installation instructions on the official website. The documentation has improved significantly and now provides clear, straightforward steps.
# Run the official installer
curl -fsSL https://openclaw.ai/install.sh | bash
After installation, an onboarding process should guide you through the setup. If it doesn’t start automatically, you can run it manually.
openclaw onboard
This process will help you connect OpenClaw to your preferred messaging platform and select an AI model.
Connecting Platforms and Choosing Models
Platforms: Telegram is the Way to Go
I have OpenClaw connected to both Telegram and WhatsApp.
- WhatsApp: The setup is finicky if you want it to appear as a separate contact. I wouldn’t recommend it.
- Telegram: This is much easier to set up and provides a superior user experience. The process involves messaging the
BotFatheron Telegram, which gives you an API token. You then paste this token into the OpenClaw onboarding prompt.
Models: The Power of Claude Opus
OpenClaw’s magic lies in its personality, and no model replicates this better than Anthropic’s own. While you can run it on various models like Sonnet 4.5, Minimax, GLM 4.7, or even GPT-5.2, I strongly recommend Claude Opus 4.5.
Many users run it using their personal Claude subscription, but this has a major drawback: OpenClaw consumes a vast amount of tokens. If you use your subscription for coding or other tasks, you’ll burn through your token limit very quickly.
A much better, and cheaper, solution is to use the Anti-gravity option during onboarding. Anti-gravity, Google’s AI IDE, offers a very generous free tier of Claude Opus 4.5. I’ve been using this and have only hit the rate limits a few times, far less frequently than when I used my personal subscription. It’s a fantastic, free, and generous option for running Opus 4.5.
The Power of Local AI: What OpenClaw Can Do
Once set up, you can message your bot on Telegram. Its real power comes from its deep integration with your data.
Why is it better than a cloud-based tool like ChatGPT?
- Full File System Access: OpenClaw runs locally on your device or VPS, giving it complete access to the file system. This is a level of access that cloud-based agents, even with APIs, cannot replicate. It’s the difference between giving an assistant a set of keys versus letting them into the entire building.
- Advanced Memory Structure: Its memory system is built on a series of local markdown files. You can instruct it to create new files for specific topics. For example, if you’re organizing a move, it can create
move_project.mdand continuously update that file with new information. This is far more transparent and detailed than the opaque memory systems of cloud services.
Concrete Use Cases
With this deep access, OpenClaw excels at personal tasks. I’ve used it for:
- Creating packing lists.
- Reading and summarizing emails.
- Booking events on my calendar.
- Conducting research.
Because it builds such a strong knowledge base about you, its answers become highly tailored and personal. For more complex projects, I’ve had it:
- Perform product analytics by giving it an API key to make requests and setting up a cron job to deliver regular analysis.
- Calculate the cost basis of my share portfolio by granting it access to my inbox, where it cross-referenced hundreds of emails to find the necessary data.
It’s incredibly effective for tasks you might assign to a personal or executive assistant but can’t yet give to a standard chatbot due to a lack of access.
The Risks and Downsides
Security is a Major Concern
It’s crucial to acknowledge the risks. You are giving an AI full access to everything on your device. If running it locally on your MacBook, it could potentially do significant damage. Be extremely careful and mindful of this when setting it up.
High-Maintenance Software
OpenClaw is still in its early stages.
- It can be buggy and sometimes fails to respond, requiring you to SSH in and fix it.
- Updates are not automatic. You must connect to the device and run the update command yourself.
It is very much a hobbyist or high-maintenance tool at this stage.
A Glimpse into the Future
Despite the maintenance, using OpenClaw feels like a peek into the future. I believe local agents will become standard. To do meaningful work on a phone or laptop, you need to manipulate local files. While cloud sandboxes can replicate this, the experience isn’t as seamless—much like Chromebooks a few years ago, which were technically functional but lacked the smooth experience of a local device.
Playing with OpenClaw is highly recommended, not just for its current utility but for the glimpse it offers into the future of personal computing. It’s a lot of fun and has already proven incredibly useful.