OpenClaw

Verdict. I ran a personal agent on it, then moved off.

what it is

An AI assistant that runs around the clock on your own computer, reached through Telegram, iMessage or the web. It codes, manages notes and runs tasks while you’re away.

It needs very little, about 1GB of RAM, plus Node.js and a Telegram bot. Setup takes about 30 minutes.

npm install -g @openclaw/cli
openclaw onboard

The architecture has a few parts.

• Gateway. Every channel (terminal, web, Telegram, iMessage) feeds one agent.
• Heartbeat. Every 30 minutes it asks itself whether it should do something.
• Scheduled tasks. Cron jobs for overnight and recurring work.
• Memory. A daily diary, PARA-structured notes and identity files that evolve.
• Self-installing tools. The agent can add its own capabilities.

what I found

From Every’s OpenClaw camp on 2026-02-06, where users walked through their setups.

• Run it on Codex, which is officially supported. Driving it with Claude Code was described as against the terms of service, with a risk of bans.
• Start with no sensitive access and add permissions slowly. Give it its own accounts, or share specific calendars and folders with it the way you would with an assistant.
• Stay away from public marketplace skills. Malware turned up in top downloads.
• The room agreed the form factor is what people want. They weren’t sure OpenClaw would be the long-term foundation.

compared to

• NemoClaw is NVIDIA’s extension of the same ecosystem.
• Hermes Agent from Nous Research is another always-on personal agent.