OpenClaw за 300 ₽/мес: как запустить AI-агента на VPS

Summary

Anton Morev presents a step-by-step guide to deploying OpenClaw on a low-cost VPS instead of paying for a preconfigured hosting package. He argues that “boxed” OpenClaw services can cost many times more than a basic VPS and may be harder to customize or troubleshoot.

Server choice and requirements

• For an inexpensive test setup, Morev recommends starting with a VPS costing about 300 rubles per month, rather than buying a dedicated computer or a high-spec server.
• His example configuration includes 2 GB of RAM, 2 GB of swap, about 30 GB of disk space, and one CPU core. He says this should be enough to try OpenClaw, though it may feel slow, especially during installation.
• Swap extends available memory by using disk space, but it is much slower than RAM.
• A more powerful server can be rented later if actual usage shows it is needed.
• Morev advises choosing a server outside Russia to access services such as OpenAI-compatible model providers and Telegram. He also warns that an overseas server’s IP address or registration details may still reveal a Russian connection and lead to service restrictions.
• He uses Ubuntu because he is familiar with it. He mentions a hosting provider in the demonstration but does not vouch for its reliability or security.

Installation and security walkthrough

The tutorial uses prepared scripts to reduce manual terminal work. The demonstrated process includes:

1. Order an Ubuntu VPS and connect initially as root over SSH.
2. Create a Telegram bot and collect its token and chat ID.
3. Install Docker and use a file manager through an SSH tunnel to edit configuration files.
4. Configure an API-compatible language-model provider, model, Telegram settings, and OpenClaw port.
5. Harden the server by creating a non-root user, changing the default SSH port, disabling root login, setting up a firewall and Fail2ban, and enabling Telegram alerts for logins.
6. Create a 2 GB swap file, install the required JavaScript runtime and OpenClaw, and complete setup.
7. Access the OpenClaw dashboard through another SSH tunnel rather than exposing its web interface directly to the public internet.

Morev explains that an SSH tunnel lets a service remain accessible locally without exposing its port to the internet.

OpenClaw configuration and model choices

• OpenClaw can run in restricted mode, where it is confined to its working folder and must be explicitly permitted to access other areas, or in full-access (“YOLO”) mode.
• Morev favors restricted mode for a server that also hosts other projects. In the demonstration, he enables full access because the VPS is dedicated to the agent, while noting that the agent can then act freely on that machine.
• He recommends GPT-5.2 as a starting model, describing it as a balance between cost and capability. Users can later assign cheaper models to suitable tasks and use a stronger model for orchestration.
• As an example, Morev mentions using Gemma 4 26B with task-specific subagents.
• He reports that the demonstration used around 800 of 50,000 available credits for basic testing, including a greeting and a request to list the server’s root directory.

The video is primarily a practical installation and setup tutorial, not a formal product review. It also discusses the trade-offs between convenience, cost, customization, security, and agent permissions.

Main speaker and sources

• Main speaker: Anton Morev, associated with the Galera Moreva channel.
• Other sources and tools discussed: OpenClaw, Telegram/BotFather, SSH, Docker, Fail2ban, and the model/API provider used in the demonstration.
• Morev also refers viewers to his prepared installation scripts and an earlier video explaining OpenClaw.