AI agents on your own server
AI agents do their best work when they can keep running: answering you on Telegram at 3 a.m., finishing a long coding task after you close your laptop, or working through a schedule. A small always-on server is the natural home for them, and keeping them off your own computer means they can only touch what you put on the server.
The models themselves run at the provider (Anthropic, OpenAI, OpenRouter and others), so the server only runs the agent. You bring your own account or API key.
The one-click apps
| App | What it is | You sign in with |
|---|---|---|
| OpenClaw | Personal AI agent you talk to through chat apps | Your model provider's API key or account |
| Hermes Agent | Open-source agent by Nous Research, with memory and messaging | Your model provider's API key |
| Claude Code | Anthropic's coding agent in the terminal | Your Claude account (Pro, Max, Team, Enterprise) or Console |
| Codex CLI | OpenAI's coding agent in the terminal | ChatGPT or an OpenAI API key |
All four start at Box Plus (2 GB RAM, 2 shared cores), $2.25 a month. For several agents at once, or big code bases, take Box Max ($4.00 a month).
Why a separate server
- Always on. Agents keep working and stay reachable while your computer sleeps.
- Contained. An agent that runs commands can only reach what is on its server, not your personal files.
- Easy to reset. Take a snapshot before you let an agent loose; roll back from the client area if it makes a mess.
Good to know
- Agents' own control panels stay on
127.0.0.1; reach them with an SSH tunnel rather than opening them to the internet. - Keep API keys on the server only, and set spending limits at your model provider.
- Our Acceptable Use Policy applies to what agents do too: no spam, scraping that ignores site rules, or attacks.
- natbox is not affiliated with Anthropic, OpenAI, Nous Research or the OpenClaw project. We install their official releases.
Questions
Do I need a GPU?
No. These agents call models over the internet; the server only runs the agent program.
Can I run a local model instead?
Only very small ones, slowly: our servers have no GPU and at most 4 GB RAM. For real work, use a hosted model.
Can I run more than one agent?
Yes, if they fit in memory. Box Max is the comfortable choice for two or three.
Also see: VS Code in the browser, self-hosting on a small VPS, all one-click apps.