# Permission controls: keeping AI edits safe

> Every Atlas tool call is permission-gated against allow, ask, and deny rules before it runs, and every edit is shown as a unified diff before it is written.

Atlas keeps AI edits safe by gating every tool call against allow, ask, and deny rules before it runs. Each file edit computes a unified diff you approve before anything is written, and changes are snapshotted as git patches so they can be rolled back. In 2026 that means automated work stays auditable and under your control.

## Key takeaways

- Every tool call is evaluated against allow, ask, and deny rules.
- Edits compute a unified diff you approve before they are written.
- A read-only plan agent proposes changes before any edits land.
- Git-patch snapshots let you roll back a change.

## Permission modes

Before any tool runs, Atlas evaluates it against your rules and resolves to allow, ask, or deny. Unmatched actions default to ask, so nothing surprising runs without a prompt.

This permission model applies to every tool call, from editing a file to running a shell command. You can broaden rules for trusted actions and tighten them for risky ones, keeping a clear boundary around what the agent may do automatically versus what needs your confirmation.

## Plan before edits

Atlas drafts a plan in a read-only plan agent and asks before switching to a build agent to implement. You see the shape of a change before it touches your tree.

Separating planning from editing means you can refine the approach while no files are at risk. Once you approve, Atlas implements the plan, and each edit still passes through the diff-and-permission step. Planning and permissions reinforce each other.

## Diffs and rollback

Each file edit computes a unified diff that is surfaced for approval before writing. Atlas also snapshots changes as git patches, so a change can be diffed and rolled back.

Reviewing a real diff before it lands is the core safety primitive: you approve exactly what changes. If something is wrong after the fact, the snapshot lets you restore. Together, permissions, plans, diffs, and snapshots make automated edits auditable end to end.

## FAQ

### How does Atlas keep AI edits safe?

Every tool call is permission-gated against allow, ask, and deny rules, and every edit is shown as a unified diff before it is written.

### Can I control what the agent runs automatically?

Yes. You set rules that resolve each action to allow, ask, or deny; unmatched actions default to ask.

### Does Atlas plan before editing?

Yes. Atlas drafts a plan in a read-only plan agent and asks before switching to a build agent to implement.

### Can I undo an Atlas change?

Yes. Atlas snapshots file changes as git patches, so edits can be diffed and rolled back.

---

Canonical HTML: https://runatlas.sh/resources/features/permission-controls
Source of truth: aeo_pages row `/resources/features/permission-controls` (segment: Platform) (this file is generated from it, never hand-edited).
Licence: Atlas is proprietary with a free core. It is not open source and there is no public source repository.