Tools and safety¶
The agent can search the web, read pages and run commands on the machine. Everything is built so that you stay in control: a button press for anything that changes something, and a sandbox for anything that runs.
Only the admin's private chat gets tools. Group members get plain conversation.
Web¶
- Search. Uses Brave if you set
agent.search.brave_key, DuckDuckGo otherwise. Setagent.search.mode: offto give the agent no search. - Reading pages. Only
httpandhttps, never private or local addresses, with size and time limits. A page you gave in your message, or one that came from search results, is read without asking. A page address the agent made up itself needs your approval.
Host commands¶
The agent can run commands as the node's user, with no shell, inside a sandbox. Each command falls into one of three classes. The model cannot choose the class.
| Class | Examples | What happens |
|---|---|---|
| Runs by itself | df, du, free, ls, ps, docker ps |
No button. |
| Needs a press | cat, grep, rm, kill, docker rm, any unknown program, bash -c |
A card shows what it will do in plain words, with the agent's reason marked as the agent's own words. You press Approve or Deny. |
| Forbidden | sudo, su, docker run, docker exec, crontab, systemctl --user start |
Never runs. |
For several steps, the agent posts a plan, and one Run button approves exactly the listed commands. Each step with a command gets ✅ when the command succeeds and ❌ when it fails. Long commands continue in the background with a live card.
The sandbox keeps the command away from the node's own data, secret files, .env files and a list of common secret stores. Add more with sandbox.deny. It cannot stop kill from reaching your other processes.
If the machine cannot offer a sandbox (not Linux, running as root, a kernel without Landlock), host commands are off and /status says why.
The kill switch is /commands off. /commands on turns them back on.
Confirmation buttons¶
By default, every action that changes something needs a button press. A decided question folds to one line such as "Approved, 21:09". A question that nobody answers expires.
Your reminders and scheduled tasks are the exception. In a clean session a reminder is set at once, and you see it with a Delete button; deleting one of your reminders or tasks also happens at once, and you get one short line saying it is off (how to stop a reminder). A tainted session still asks for both.
If you change your request and the agent asks the same kind of question again (for example, a reminder for another time), the old question is replaced: its buttons go away and it folds to "Replaced by a newer question", so only the current one can be pressed.
Tainted sessions¶
When the conversation reads something an outsider wrote (a web page, a search result, a forwarded message, a quote from another chat), the session is tainted. In a tainted session the agent might be reading instructions planted by someone else, so it becomes stricter. Any action that changes something, any message to someone else, any memory write and any script that takes input needs your press, even ones that would have run by themselves.
Send /new to clear it. It also clears by itself after 6 hours of silence.
Secrets¶
The model never sees a secret. Secrets are redacted from logs and the tool-call journal before anything is written. Keep them in env: variables or the encrypted store, and never type one in chat. See Configure.