The agent reads your project and runs analysis commands freely, because that is most of the work. It also publishes its reports, with the charts, scripts and other results it made under output/, to the project your folder is linked to without asking, and tells you each time with the link: the people with access to that project can read them, which can include guests from outside your organisation. It stops and asks before it does anything else that changes your environment or shares your data. Before a job that would run for a long time or cost much, the agent proposes it in the conversation and waits for your answer. When your question needs data from a service you use, it proposes connecting it and walks you through it; the connection itself still waits for your approval.

Questions

When a choice is yours to make, the agent asks a question with options. Use ↑ ↓ to choose, Enter to confirm, or type your own answer. Several questions appear as tabs; review them on the last tab before you submit. A key or password for a connection is asked for in a hidden field instead: it shows dots, never the key, and the review shows it as “hidden”. See Keys and passwords.
A question in the TUI asking whether to attach the analysis code, executed notebook and aggregate results to the report, with the options Approve evidence attachments, Report text only and Type your own answer

The agent asks which evidence to attach to a report

Permission prompts

When a command or tool needs permission, the terminal shows the exact command and these choices:
  • Allow once — run it this time only.
  • Always allow — allow this exact thing from now on: the same command or the same folder, so a different command still asks.
  • Allow all service — for a tool of a connected service (for example Allow all gmail-work): allow every tool of that service for the rest of this session, so setting up an account asks once, not once per tool.
  • Allow all — allow everything for the rest of this session, including requests already waiting, and never ask again for it. In AIUS Light’s remote view of a conversation the same answer appears as Allow all.
  • Reject — refuse it; nothing runs, and the agent looks for another way.
Anything a rule explicitly denies still asks or refuses, even under Allow all. The sidebar’s Permissions & security line shows what Aius may do without asking right now and, in a folder linked to a project, what syncs from it; select it for the list and a link to your organisation’s Permissions & security tab in AIUS Light. The command palette also offers the wider version — Auto-approve all permissions (dangerous), which approves everything for every session the terminal runs until you turn it off or the terminal exits. aius --auto starts the terminal with it already on.
aius --auto and the all-sessions bypass approve every request that is not explicitly denied, including package installs and publishing. Use them only in a disposable environment.
When the folder is linked to a project, the organisation it belongs to can set a conversation permission policy that decides first: its denials always refuse, its allow-list runs without asking, its ask rules always ask you, and it can keep the allow-all bypass off entirely. To be asked before every report as well, tell the agent “Always ask before publishing”. It saves that in your settings, as "artifact_publish": "ask" and "aius_platform_report_publish": "ask" under permission. Aius reads its settings only from ~/.config/aius-agent/aius.json and, in a project, from aius.json or .aius/aius.json. Settings written for other coding agents never change what Aius asks before it acts.