AI Agent Notes - (2)

Recently, I have been thinking more about how to run an AI agent safely on my own computer.

A setup I like is to run OpenClaw inside a virtual machine. The VM becomes the main agent environment. I only expose a small web port for quick app preview, and use SSH when the agent needs to work on the host side.

In my case, the host can be Windows or WSL2. I add that machine as an SSH host, so the agent inside the VM can edit code there, run commands there, or manipulate Docker there, without living directly in the host environment.

This gives one more guardrail layer between the OpenClaw runtime and my main working system.

A simple sketch is:

1
2
3
4
5
6
7
8
9
10
11
12
13
14

[Browser]
    ^
    | web port
    |
[VM: OpenClaw]
- isolated runtime
- guardrail layer
    |
    | SSH
    v
[Windows / WSL2]
- code
- Docker
- SSH host

Another point is AI identity.

I think it is better to give the agent a separate email account, a separate GitHub account, and a separate browser profile. That reduces the risk of accidental deletion, wrong login state, or mixing personal data with agent actions.

For me, this kind of setup is not about making the agent stronger. It is about making it safer, cleaner, and easier to trust.

AI, OpenClaw, agent, security — Jul 1, 2026