AI can take real action. You decide the limits.
Before work begins, decide what an agent may read, change or send, and which actions must stop and wait for you. Bellfly keeps those choices with the work.
Start with the questions that matter to your project.
A useful boundary names the information and the action, not just the tool.
What can it read?
Choose the project materials and sources needed for the task. Access to one folder is not access to all files on your computer.
What can it change?
Set the allowed files and output locations. Keep originals when a transformed copy is what you need.
When must it ask?
Set approval points for publishing, deleting, purchasing, sending messages or increasing costs and access.
Where do credentials belong?
Use the operating system’s secure credential storage or the connected tool’s authentication. Keep passwords and tokens out of prompts, project notes and reports.
Could a failed action be repeated?
When an outside action has an uncertain result, Bellfly checks what actually happened before trying it again. Stopping work does not undo an earlier action.
Tell us when a boundary does not behave as expected.
Share the affected version, what you expected and a minimal example with sensitive content removed. Do not send passwords, tokens or private project files.