Guide
How do you set up a self-hosted AI agent safely?
Treat the agent like a new employee with admin access: run it in a sandbox, give it only the keys it needs, require approval for risky actions, and limit who can message it. Then choose whether to build this yourself or have someone do it.
What goes wrong
An agent that can read files, use accounts and browse the web is only as safe as its configuration. A loose setup can expose keys or act on instructions from anyone who can reach it.
Controls to put in place
These apply whatever agent software you choose.
- Sandbox the agent so it cannot reach files and processes it does not need.
- Keep secrets in an encrypted key store, never in prompts or plain config files.
- Add permission gates so destructive or external actions need approval.
- Allow only paired, known senders on each messaging channel.
- Log what the agent does, and review it.
Decide where it runs
A machine you own keeps data local but needs upkeep; a hosted instance is quicker but your data sits on someone else’s infrastructure. Pick based on what the agent will touch.
A worked example: SetClaude
SetClaude is a setup service for the OpenClaw agent gateway. Its hardening covers sandboxing, an encrypted key vault, permission gates, and DM pairing with an allowlist, and it can run hosted by Cuanto Labs or on the client’s own Mac Mini.
The process is a 30-minute call, one to two days of configuration, a live walkthrough, then ongoing support. Channels include WhatsApp, Telegram and Discord, with Playwright and Chrome for browser automation.
More on the example: SetClaude.