Claude auto mode: what it is and how to turn it off
2026-10-23 · 4 min read
Claude auto mode is a permission setting where Claude skips the "approve this step?" prompts and a separate automated check screens each action, blocking the risky ones. Starting next week it becomes the default in the Claude web, desktop and mobile apps on Pro, Max and Team plans, and you can turn it off by picking Manual in the mode selector in the chat box.
Anthropic announced the change on Thursday. @claudeai wrote that instead of asking you to approve each step, "an automated check blocks risky actions," and that you can switch back to manual anytime. The post gave no exact date beyond "next week."
Claude Code users have seen this already. Auto mode became the default for new Claude Code sessions on paid plans in mid-August, as TechCrunch reported. This week's change brings the same idea to the regular Claude apps, where the agent work that used to live in Cowork now happens inside ordinary chat. We covered that merge in Claude now picks chat or agent for you, and at the time approval was still the default. Next week, on most paid plans, it won't be.
What auto mode does
Anthropic's Cowork help page lists three modes, and the selector sits in the chat box:
| Mode | What happens | Default for |
|---|---|---|
| Manual (formerly "Ask before acting") | Claude pauses on actions and you click Allow or Deny | Enterprise |
| Auto | Claude screens each action for risks such as data exfiltration or prompt injection, and blocks unsafe ones | Pro, Max, Team |
| Skip (formerly "Act without asking") | Nothing checks actions automatically | None listed |
Connectors are where you'll feel the difference. In auto mode, a connector you've marked Blocked stays blocked. Read-only tools, such as looking up a calendar or reading a file, are approved without asking. For tools that write or delete, Claude makes the call itself. In Manual, any connector set to "Needs approval" still asks you first.
The help page says auto mode covers your connectors, plugins, the built-in browser, Claude in Chrome, and some Cowork actions such as fetching websites. When the check blocks something, Claude tries a safer route or asks you. If blocks keep happening, it goes back to asking at every step. Permanently deleting files always needs an explicit Allow, whatever mode you're in.
How to turn auto mode off
In the Claude apps, open the mode selector in the chat box and pick Manual. You can change it in the middle of a task, too.
On Team and Enterprise, the admin decides whether auto approval shows up for the organization at all and whether new tasks start in it. If it's been turned off, you won't see it in the selector. Admins can also require approval for every connector tool that can write, which overrides anyone's "Always allow" settings.
In Claude Code, Shift+Tab cycles through the permission modes, according to Anthropic's auto mode post.
My read
For most people this is the right default. Approval prompts lose their value once you've seen fifty of them. People stop reading and click Allow, and then the prompt protects nobody. I'd rather have a check that screens every action for exfiltration and injected instructions than a tired person pressing a button.
It can still get things wrong, though. Anthropic's own March post on auto mode says it can let some risky actions through when intent is unclear, and that it sometimes blocks harmless ones. The help page also says plainly that no mode replaces your judgment, and it suggests staying close or switching to Manual for high-stakes work. I'd take that advice literally.
Here's the split I'd use for a small business:
- Leave auto mode on for research, drafting, summaries and anything read-only.
- Before next week, open Customize > Connectors and set anything that sends, posts or pays to Needs approval or Blocked. Auto mode respects Blocked, and Manual respects Needs approval.
- Switch to Manual for a task that touches money, customers or anything you can't undo.
Other agents are going the same way. Meta routes every outbound Muse action past a second AI, and Anthropic's new Claude Duet agent still waits for your OK before it makes a payment. The automated reviewer handles the routine steps, and you get asked about the ones that move money or can't be undone.
What it means for your business
If your team uses Claude on a paid plan, its behavior changes next week whether anyone reads the announcement or not. Spend ten minutes this week listing which connectors can write to which systems, and decide which of those always need a person. If you'd like help drawing that line across your actual workflows, our free process audit starts with exactly that map.