Ask, Plan & Auto modes

The AI copilot only ever acts by running shell commands — each one signed by your browser and executed on the server through the secure agent. What changes between sessions is how much it's allowed to do on its own. That's the mode. You pick one at the top of the chat before (or during) a session, and you can switch at any time — open a server in Servor and switch to its AI tab to try it.

Mode	What it does	Best for
Ask	Read-only. Advises and explains, never changes anything.	Learning, second opinions, safe investigation.
Plan	Proposes a full, ordered plan you approve once, then runs the steps in order.	Multi-step changes you want to review as a whole before anything runs.
Auto	Runs commands on its own — except destructive ones, which still need your approval.	Routine, low-risk work you don't want to babysit.

Ask — read-only advice

In Ask mode the copilot won't change anything on your server. It can still read context — check a service, tail a log, look at disk usage — but it uses that only to answer your question and suggest what to do next. Nothing it proposes gets applied.

Use Ask when you want to:

• Understand a problem before touching anything ("why is my site slow?").
• Get a recommended fix you'll apply yourself in the web terminal.
• Learn how something on the server is configured.

Think of it as a knowledgeable colleague looking over your shoulder — full of advice, hands in their pockets.

Plan — approve the whole plan, once

Plan is the careful middle ground, and a good default. The copilot works out everything that needs to happen and proposes a complete, ordered plan — every step, each with its command and a risk level. You review the whole thing up front and approve once. Servor then runs the steps in order on your behalf, signing each command in your browser as it goes.

The safety net stays on while it runs: a destructive step still pauses for a per-step confirmation. A read-only inspection step (like systemctl status) that returns a non-zero code is treated as informational, not a failure — the plan keeps going. You watch each step turn from pending to running to done as the plan executes.

And if a step fails for real, the plan doesn't stop right away: the copilot tries to repair itself — it diagnoses, tries another approach, and re-runs the step, up to five attempts before giving up. You don't have to do anything; the step shows "self-repairing…" while it works. The one exception to "nothing to do": if a repair command is destructive, it asks for your approval, like everywhere else.

Once the plan finishes, the copilot picks the thread back up: it reads the result of each command and tells you where things stand — what it found, what worked, what's left. The plan itself stays in the conversation (it collapses to a summary once done), rather than being pinned above the composer.

Each step records every command it ran — its own and any tried while repairing. They show live under the step, collapse once the step is done, and stay reviewable with one click on the step: so you can check afterwards exactly what was executed on the server.

Use Plan when you're making real changes but want to see the shape of the whole operation before it starts — a multi-step install, a hardening pass, anything on a production box you care about.

Auto — hands-off, with a safety net

In Auto mode the copilot runs safe commands on its own and keeps going, showing you results as it works. It's also the dynamic mode: it reads the result of each command and adapts what comes next — if something fails, it tries another approach instead of stopping cold. It's the fastest way through routine, repetitive work. The safety net stays on: anything risky or destructive still pauses for your approval, exactly like in Plan. Auto only removes the prompt for the low-risk commands.

Use Auto for:

• Routine health checks and diagnostics.
• Multi-step read-heavy work where approving each step would just be clicking "yes".
• Tasks you've done before and trust the copilot to handle.

Switching modes mid-session

You can change modes at any point in a conversation. A common flow:

How modes relate to roles

The mode controls the copilot; your role controls whether you can use the copilot at all. Viewers can't run commands, so they can't start a session in any mode. Members, admins and owners can. See Roles and permissions.

See also