What Claude Code's Auto Mode Actually Blocks for You

Share

Auto Mode has a classifier model check each action before it runs. By default it blocks risky steps like running downloaded scripts right away, production deploys and force pushes. If it keeps getting blocked, it asks you again. Here is how it compares with turning all permissions off.

Machine translation. Read the original

SourceAnthropicReported Nov 15, 2026Not yet tested

Claude Code's Auto Mode hands the permission prompts you normally click "Allow" for to a separate classifier model. The model reviews each action before it runs, lets safe ones through, and blocks dangerous ones. According to Anthropic's official docs, what it saves is the time you spend watching the screen and clicking confirm. The docs also state plainly that it does not guarantee safety.

What It Blocks for You

The default blocks listed in the official docs are all actions that are hard to undo:

Blocked by default Example
Downloading and running right away `curl ...
Sending sensitive data outside Keys, internal files
Changing production Production deploys, database migrations
Mass-deleting cloud resources Wiping a whole storage bucket in one go
Rewriting Git history Force pushes, git reset --hard

Everyday actions like reading and writing files in the project folder and running tests are allowed right away.

There is also a safety net. If it is blocked 3 times in a row, or 20 times total in one session, Auto Mode pauses and goes back to asking you at every step. When Claude keeps trying risky moves, the decision goes back to you.

How It Differs from Turning Everything Off

Some people find the prompts annoying and just turn all permissions off. The two are very different:

Auto Mode Everything allowed (bypassPermissions)
Is anything checked before it runs? Yes, the classifier model checks each action No
Risky actions Blocked by default Run directly
Where the official docs suggest using it Everyday development Only in isolated containers or virtual machines

So in the end, people who want to save effort without risking accidents mostly keep Auto Mode on.

Cost

Reviews run on the classifier model. According to the official docs, in a session connected directly to Anthropic, the server can do the review together with the normal request. When the server does not do it (common with third-party gateways or proxies), Claude Code sends its own classifier request instead. On accounts that bill for those requests, it shows a billing notice. Some user reports blame unusual bills on Auto Mode. There is no bill or setting to check that against, so these reports are not verified yet. If you suspect it is using up usage, check where the usage went with /usage first.

Notes

  • Starting with version 2.1.283, interactive sessions in the terminal and VS Code start in Auto Mode by default. This is the same for all plans.
  • To make it your default mode, put defaultMode: "auto" in the user-level ~/.claude/settings.json. It does not work if you put it in project settings.
  • The classifier model can also get it wrong. Check important actions yourself before they run, and keep backups.
  • For scheduled jobs and other runs that nobody watches, think hard about what you allow them to do first.

Official details are in Claude Code's permission modes.

Done with this one? Mark it to track your path.Sign in
Save it for laterSign in
Questions about this piece?Ask AI about this

Comments (0)

No comments yet. Start the conversation.

Sign in to join the discussion.Sign in