PicoClaw setup: VPS, model provider and Telegram

In this Russian-language tutorial, Anton Morev walks through installing PicoClaw as a self-hosted AI agent on an empty VPS using his prepared scripts. He reports memory use below 10 MB and says he has run it on Raspberry Pi hardware, an Android phone and servers. That figure refers to the agent software: PicoClaw sends requests to an external LLM provider rather than performing inference itself. This setup therefore does not demonstrate offline model use.

The tutorial starts with SSH access and basic server security, including a separate user, disabling root access, fail2ban and a firewall. Morev then installs PicoClaw, opens its web interface, enters an API key and selects a default model through his own hosted provider, Wormsoft. He demonstrates restricting interface access to a specific IP or using an SSH tunnel to reach localhost on port 18800.

Configuration covers workspace restrictions, a 50-call tool limit, session separation by user and channel, MCP connections and Telegram access. Morev describes using MCP to reach his computer's browser and retrieve information from an authenticated session. He disables heartbeat checks and prefers scheduled tasks.

He describes using PicoClaw to prototype a travel knowledge base with Obsidian before building a separate custom application, and reports using it to help repair server configurations. He treats PicoClaw as a way to test ideas before building dedicated software. He also reports difficulty with custom scripts and hooks, and says he did not find built-in orchestration across separate Telegram chats.