Skip to content
eWiki The practical AI reference

AI Tool Glossary

What Are Claude Code Permission Modes and How Do They Control Tool Approval?

Isometric white network of raised paths and blocks with several dark gateway structures accented in blue.

Claude Code permission modes set a session’s baseline for which actions proceed automatically, require approval, or are denied. Six documented modes support workflows ranging from closely supervised work to isolated automation, while explicit permission rules provide additional control.

How Claude Code Permission Modes Affect Approvals

A Claude Code permission mode establishes the baseline approval behavior for a session. Under the Manual (default) baseline, reads and Grep within working or additional directories can proceed without approval, while most file modifications, Bash commands, web fetches, and web searches prompt the user.

The selected mode is only one layer of control. Fine-grained allow, ask, and deny rules can change how particular tools or actions are handled. This distinction matters when choosing a workflow: a more permissive mode does not necessarily authorize every action, and a restrictive mode may still permit explicitly approved tools. Users who need to inspect sensitive changes can retain interactive oversight, while those automating repeatable work can configure more predictable behavior.

Three metal toggle switches, a large blue dial, calipers, rulers, a pen, and a closed book arranged on a work surface.

The Six Permission Modes Explained

Claude Code documents six permission modes:

  • Manual (default) allows reads without asking but generally requests approval for file edits, shell commands, and network actions. The documentation presents it as appropriate for sensitive work and reviewing actions.
  • acceptEdits permits reads, file edits, and common filesystem commands without permission prompts. It reduces interruptions when editing is expected, while other controls can still apply.
  • plan supports codebase exploration before changes. It permits reads and, where auto mode is available, commands approved by the classifier.
  • auto uses a separate classifier to review actions instead of asking the user to approve each one. This changes who performs the action-level review; it does not eliminate the surrounding permission system.
  • dontAsk permits reads inside working directories and pre-approved tools. Actions that would otherwise trigger a prompt are denied, subject to the documented plugin-mod behavior. That predictable denial model is intended for locked-down scripts and CI where no user is available to respond.
  • bypassPermissions permits actions without the normal checks and is documented only for isolated containers or virtual machines.

Choose according to the task rather than convenience alone: Manual emphasizes direct review, plan emphasizes exploration, dontAsk favors noninteractive execution with denials, and bypassPermissions is limited to isolated environments.

Two people stand at a large white table, examining a hand-drawn flowchart on paper in a bright, minimalist room.

Permission Rules, Mode Switching, and Common Questions

Permission rules layer allow, ask, and deny decisions over the active mode. A matching deny rule blocks an action in every mode, including bypassPermissions. Some actions also are not automatically approved in any mode, including tools covered by explicit ask rules, tools requiring user interaction, and recognized critical-path removal operations.

A running session’s mode can be changed with Shift+Tab in the Claude Code CLI or with mode controls in supported graphical interfaces. The initial mode may vary because command-line flags, settings, built-in defaults, interface, plan, provider, or organization policy can affect it, so users should verify the active mode rather than assume it.

Approval duration also differs by action. For Bash commands and WebFetch domains, selecting a persistent “don’t ask again” option can save a repository-scoped rule. Approval for file modification lasts only through the current session. These scopes make the effective permission behavior a combination of the mode, saved rules, session approvals, and applicable policy.

Conclusion

The practical choice is a balance among oversight, automation, and predictable denial. Repeated prompts can interrupt longer work, but reducing prompts also changes how actions are reviewed. Select Manual for close supervision, plan for exploration, or dontAsk for locked-down noninteractive work; reserve bypassPermissions for its documented isolated environments. Before starting your next Claude Code task, identify whether it needs interactive review, exploration, or unattended execution, choose that workflow’s baseline, and inspect the explicit permission rules that will apply.

Disclosures and limitations

  • This article was prepared with AI assistance using the supplied Claude Code documentation and related source materials.
  • References to approval fatigue reflect supplied public commentary; product behavior and configuration details are attributed to the supplied documentation.

Sources

Improve this guide

Report a problem or suggest the next question

Submissions enter a private editorial queue. They are reviewed and deduplicated before any public update or new page is created.