Can an AI help me manage my Linux server without doing something stupid?
A sysadmin on the line who proposes the next command, explains it, and waits for you to approve.
What you are running into
Disk full at 2 a.m. You know the answer is somewhere in du, journalctl and apt clean, but not the exact flags, and a wrong rm is not a thing you can undo. Pasting terminal output into a chat window works until you paste a token you should not have, or copy back a command that was confidently wrong.
Agents that run commands on their own solve the typing and create a worse problem: a model with a shell and no adult in the room.
How SSH Guru handles it
Propose, never execute
The Guru watches the terminal and suggests the next command as a card with a reason and what to expect. It cannot run it. You click approve, edit it first, or skip.
Risk tiers decided by code
Read-only commands are tier 0 and, on Pro, can run themselves if you allow it. Anything that changes state is tier 1 and needs a click. Destructive commands are tier 2 and make you type the hostname. The model does not assign tiers; fixed rules do.
Secrets scrubbed first
Command output passes through a deterministic scanner before any model sees it. Tokens, keys and passwords become placeholders. Optionally IPs and hostnames too.
What to do, step by step
- Create an account
The free plan includes house models and 30 Guru messages a day. On Pro you can bring your own brain: local Ollama, an OpenAI, Anthropic, OpenRouter or Groq key, or your ChatGPT subscription.
- Add a server and connect
The SSH client runs in your browser; your key is encrypted with your passphrase.
- Describe the problem in the chat
"web-01 is almost out of disk, find what is eating it." The Guru starts with read-only commands.
- Approve, edit or skip each card
Every card shows the command, why, and the tier. Edit it and the tier is recomputed. Type the hostname for anything destructive.
- Keep the session
Chat and terminal history stay together, so next month you can see exactly what fixed it. On Pro the same session becomes a runbook you can replay.
Questions
Which models can it use?
House models on the free tier. On Pro, your own API key for OpenAI, Anthropic, OpenRouter or Groq, a local Ollama, or your ChatGPT subscription through the same sign-in the Codex CLI uses.
Does the model see my whole terminal?
It sees the output of the commands it proposed, after scrubbing. What you type by hand stays out unless you share it.
Can it be tricked by something in a log file?
Prompt injection is real and the assistant is told to treat output as data. The thing that actually stops a bad command is that nothing runs without your click in the browser.