
Permission modes: how much a coding agent may do unasked
In short: Permission modes set which actions a coding agent may take without asking, from reviewing every step to skipping checks inside a sandbox.
Permission modes are settings in coding agents that decide which actions the agent may take without asking you first. Claude Code's documentation, read in November 2026, lists six: default (labelled Manual, asks before edits and commands), acceptEdits, plan, auto (a classifier model reviews actions), dontAsk and bypassPermissions. Allow, ask and deny rules fine-tune specific tools on top of the mode.
Permission modes in plain words
Think of how long a leash you give the agent: a short one that checks every step, a longer one for routine edits, or none at all in a fenced yard.
How it works
Every proposed action, such as an edit, a shell command or a web fetch, is checked against the current mode and the rules in your settings files. The permissions reference evaluates rules in a fixed order: deny, then ask, then allow. A deny rule blocks in every mode, including bypassPermissions. Organisations can deploy managed settings that user and project settings cannot override.
The same documentation says recent versions start interactive terminal sessions in auto mode, so check which mode you are in before assuming every step will ask.
Example
.claude/settings.json allowing routine commands while blocking risky ones:
{
"permissions": {
"allow": ["Bash(pnpm test *)", "Bash(pnpm lint *)", "Bash(git diff *)"],
"deny": ["Read(./.env)", "Bash(curl *)", "Bash(git push *)"]
}
}
The reference warns that a Bash rule matches command text and is no security boundary: Bash(curl *) does not stop /usr/bin/curl.
Where you'll meet it
- Claude Code's Shift+Tab mode cycle and settings files
- Approval settings in other coding agents
- Managed settings in organisations
- CI runs, where
dontAskdenies anything that is not pre-approved
What it changes for you
For beginners, use Manual mode for sensitive or unfamiliar code, and acceptEdits once you review diffs before committing. My view: an allow-list for safe commands removes most prompt fatigue. The documentation limits bypassPermissions to isolated environments such as containers or VMs. For security teams, managed settings can disable bypass and auto modes outright.
Permission modes vs sandboxing
Permissions decide which tools the agent may use and what it may access. A sandbox enforces filesystem and network limits on shell commands at the operating-system level. The Claude Code documentation calls them complementary layers.
Common mistake
My view: switching to bypass mode on a laptop holding production credentials because the prompts got annoying. Fix the annoyance with an allow-list, or move the work into a container. Do not remove both brakes at once.
Related terms
Agent sandbox · Plan mode · Claude Code hooks · Human-in-the-loop · Claude Code
From the archive: Agent approvals need independent policy checks
Cut prompt fatigue with an allow-list or a sandbox, and never by removing both brakes at once.


