Ask and Agent
The assistant, its two modes, the approval gate and how secrets are kept out.
Providers
Add a provider in Settings > AI assistant. OpenAI-compatible endpoints and Anthropic are supported. Provider keys are stored in the vault, never in the settings file.
Modes
Ask answers questions with the visible terminal as context.
Agent can also run commands in that terminal, and read and change files over SFTP.
In Ask mode the assistant can request a switch to Agent when a request needs a tool. If you approve, the request is retried in Agent mode and the choice is remembered. If you decline, the assistant answers without the tool.
Risk and approval
Every command is classified before it runs:
| Level | Meaning |
|---|---|
| Low | Read-only inspection with no side effects. |
| Medium | Changes files or configuration with a bounded effect. |
| High | Deletes data, changes permissions or privileges, stops services, runs unseen code, or is otherwise hard to undo. |
| Critical | Can destroy the machine, its data or the remote access to it. |
Low-risk actions run; the rest ask for your approval first.
The approval gate is a safety control, not a sandbox. Do not point an agent at a host where a wrong command is unacceptable.
What you see in the terminal
A command the assistant runs is shown as the command itself, with an accent mark in the terminal's left margin. Motty hides its capture wrapper and marker from the screen and from its own command history. The remote shell's own history is outside Motty's control.
Secrets
Secrets are redacted before terminal text reaches a model. Redaction is on by default.