Claude Code auto mode: wetin the new default mean
Auto mode na default for Claude Code v2.1.283+ terminal and VS Code sessions. See permission modes and which one fit server wey you no dey watch.
Wetin auto mode dey change on 14 August 2026
Claude Code auto mode dey run tool calls without routine permission prompts, and e dey use separate classifier model to review actions. Reads and edits for working directory outside protected paths dey skip the classifier. From 14 August 2026, na the mode new sessions dey start with for Pro, Max, and Team plans. As of 27 October 2026, the documentation don explain am further: with Claude Code v2.1.283 or later, auto mode na the built-in starting mode for interactive terminal and VS Code sessions on every plan and provider. You fit switch modes anytime, and any default wey you don already set for yourself no go change.
The documentation now talk the change like this:
With Claude Code v2.1.283 or later, auto mode na the built-in starting permission mode for interactive terminal and VS Code sessions. For earlier versions, na the built-in starting permission mode only for Pro, Max, Ultra, and Team plans.
Two rules for the same page matter pass the version number. Any defaultMode wey you set inside your own ~/.claude/settings.json go remain after the change: Claude Code fit ask you once whether to switch am to auto mode, and e go keep your setting if you decline. Any default wey your organization deploy through managed settings go remain too. The announcement post talk say auto mode go remain opt-in on Enterprise plans and the Claude API at first, and an update to am on 22 October 2026 say Enterprise admins go receive email before auto mode become the default for their organization. From v2.1.283, the documentation no longer make that exception for terminal and VS Code sessions, although with v2.1.285 or later, any organization wey get the HIPAA configuration applied still dey start in Manual mode.
If you dey run Claude Code for VPS (virtual private server), this change matter pass how e matter for laptop. Permission prompt na checkpoint wey person for keyboard need handle. For remote box, you often no dey there, so the mode wey session start with na the mode e go stay on for hours. Non-interactive claude -p runs follow another rule: with v2.1.302 or later, dem dey start in default unless you set a mode, including when you use third-party provider or turn telemetry off. As of 31 October 2026, the current release na v2.1.316, and the permission-modes page document this behaviour. Before v2.1.302, those two cases fit start in auto, so run claude --version for the box if e fit still dey use older release.
Claude Code permission modes, from the one wey get the most oversight reach the one wey get the least
There are six modes. The name for the beginning of each line na the value wey you write for settings or pass to --permission-mode. Claude Code itself dey enforce all six, no be the model, because deciding wetin tool call fit do na the work of the harness around the model.
default: Claude go ask before most actions wey edit files, run shell commands, or reach the network. Reads inside your working directory still go run without prompt. The CLI (command line interface) labels this mode Manual and acceptsmanualas alias.plan: Claude dey read files and run commands to explore, but e no dey edit your source. Edits remain blocked until you approve the plan, except for interactive terminal sessions wey get bypass permissions available: for there, neither classifier nor prompt dey apply to planning commands.acceptEdits: file edits go run without prompt, together with filesystem commandsmkdir,touch,rm,rmdir,mv,cpandsed. This one apply only to paths inside your working directory or youradditionalDirectories. Writes to protected paths and recognized critical-path removals still go ask, andtouchgo auto-approve only with plain file paths. Every other shell command go follow the normal permission rules.auto: routine permission prompts go replace with classifier checks. Explicitaskrules still go force a prompt.dontAsk: Claude Code go auto-deny anything wey for normally ask you. Apart from calls wey a mod wey you install fit approve, file reads inside your working directories, yourallowrules, the built-in read-only Bash commands, preapproved documentation fetches, and calls wey aPreToolUsehook approves, everything go run. The session no go ever wait for input.bypassPermissions: prompts and safety checks go skip, including writes to protected paths like.gitand.claude.
Press Shift+Tab during a session to cycle from default to acceptEdits to plan. From auto, the first press go drop you to default. The status bar go show where you land, like ⏵⏵ auto mode on or a gray ⏸ manual mode on. The other modes no dey inside that cycle by default. auto go join am once your account meet the requirements. bypassPermissions go join am only when the session start with --permission-mode bypassPermissions, --dangerously-skip-permissions or --allow-dangerously-skip-permissions, or with defaultMode: "bypassPermissions" for user, --settings or managed settings. dontAsk no dey appear there, so set am with claude --permission-mode dontAsk.
Auto mode also need a recent model, and na the usual reason why e no dey appear at all. As of 27 October 2026, the documentation list Claude Opus 4.6 or later, Sonnet 4.6 or later, Haiku 5.5, or a Fable model for the Anthropic API. The list shorter for Amazon Bedrock, Google Cloud and Microsoft Foundry. E also talk say older models like Sonnet 4.5 no dey supported for any provider. If Claude Code report say auto mode no dey available, check those requirements and any disableAutoMode setting first. Anthropic fit also disable auto mode from the server side, and a session wey receive that answer go keep auto mode disabled until e end. So start a new session later instead of waiting inside the old one.
Two controls dey hold for every mode, including bypassPermissions: deny rules and explicit ask rules. Neither one apply to EndConversation while another tool still dey available. Na those controls you keep, no matter the mode wey session start with.
Auto mode classifier dey block wetin
Classifier na second model wey dey read action wey dey wait, then decide whether e match wetin you ask for. Documentation describe the work for one sentence:
Separate classifier model dey review actions before dem run, and e dey block anything wey pass wetin you request, target infrastructure wey e no recognize, or look like hostile content Claude read dey drive am.
By default, e dey block these ones, especially the categories wey server operator dey meet often:
- Downloading and executing code, like
curl | bash - Production deploys and migrations
- Force push
- Modifying shared infrastructure
- Opening tunnel or reverse shell wey make local service reachable from public internet
- Printing live credential or token inside transcript or file
By default, e dey allow:
- Local file operations inside your working directory
- Installing dependencies wey your lock files or manifests declare
- Read-only HTTP requests
- Pushing to branches for the repository wey you dey work on, including default branch, except non-default branches wey their names show say dem be deploy or publication targets. The push content still pass through the other classifier rules.
No use summary like the one above as your working source. Run claude auto-mode defaults to print the complete rule lists as JSON, then read the set wey ship with the version wey you install. /auto-mode-setup na separate command: e dey draft autoMode.environment entries from your project and recent sessions inside am, then write dem to ~/.claude/settings.json if you accept the draft. Na that command, not claude auto-mode defaults, carry the version requirements. For Pro, Max, Ultra and Team plans, /auto-mode-setup need v2.1.228 or later, or v2.1.233 or later for native Windows. For Enterprise plan or when you use Claude Console API key, e need v2.1.283 or later. Before v2.1.304, /auto-mode-setup no dey inside sessions wey no fetch feature flags, such as server wey telemetry turn off or server wey dey use third-party provider. So update first if that describe your server.
Two documented limits matter before you depend on this. First, for classifier requests wey Claude Code send by itself, classifier dey see your messages, tool calls apart from read-only lookups like file reads and searches, and your CLAUDE.md content. Tool results dey remove, so text inside file or web page wey Claude read no fit address classifier directly. Second, when classifier block action 3 times one after another or 20 times for one session, auto mode pause and Claude Code return to prompting you. If you approve the prompted action, auto mode resume. You no fit configure these thresholds. Non-interactive run wey get -p flag but no --permission-prompt-tool get nobody to prompt, and documentation clear say Claude Code no stop the run: blocked action no run, while Claude continue to work.
Na this second behaviour dey cause problem for remote box, in 2 different ways. Interactive session wey remain running and reach the limit go stop for prompt, then wait for person wey no dey look the terminal. A -p run no ever wait, so job fit finish while blocked steps remain undone. Second Claude Code session for the same box no be replacement for that person, because text wey one session send to another no ever count as your consent and e no fit answer pending permission prompt. So session wey pause for one still need human answer. To narrow wetin agent set out to do na the other half of the fix, and skill wey push agent toward the smallest change wey work dey stop session from entering the wide-ranging actions wey classifier dey stop. Where skill dey guide one task, output style dey edit the system prompt itself, so style wey ask for small, surgical changes go apply for every turn of the session, instead of only the one job wey you invoke am for.
Where modes dey for settings.json
Everything wey dey above na one object inside settings file.
{
"permissions": {
"defaultMode": "auto",
"allow": [
"Bash(npm run test *)",
"Bash(git status)"
],
"ask": [
"Bash(git push *)",
"Bash(docker compose up *)"
],
"deny": [
"Read(./.env)",
"Read(./secrets/**)",
"Bash(curl *)"
]
}
}Dem dey evaluate rules in this order: deny, then ask, then allow. The first match for that order decide the result, and a narrower rule no override a broader rule wey come earlier. A deny rule for Bash(aws *) block aws s3 ls even when you also allow that exact command, so allow rule no fit create exception from deny rule. Read and Edit deny or ask rules support gitignore negation: for one settings file deny list, Read(*.env) followed by Read(!sample.env) exempt files named sample.env from the rule wey come before am. The exception only reach earlier relative-path rules from the same settings source; e no fit override deny from another source or reopen file inside directory wey dem block as a whole.
ask na the rule type wey useful for auto mode. Auto mode remove the routine prompt, and ask rule put one back for the specific command wey you want person to approve. Your deploy command belong there. Bash(git push *) belong there too if you want checkpoint before code leave the box. Keeping credential files for deny na the other half of this, and e pair with keeping credentials away from agent reach from the beginning.
The settings files themselves, from lowest precedence:
~/.claude/settings.json: na your user settings, and dem apply for every project..claude/settings.json: na project settings, and dem dey committed to the repository..claude/settings.local.json: na your own settings for one repository, and git ignore am when Claude Code create am; add am to.gitignoreif na you create am by hand.- Managed settings, wey administrator deploy. For Linux, that file na
/etc/claude-code/managed-settings.json. Nothing fit override managed permission rule, including command line flag.
One trap here get documented cause. defaultMode: "auto" no take effect when e come from .claude/settings.json or .claude/settings.local.json, so repository no fit give itself auto mode by shipping settings file. bypassPermissions dey ignored from those two files too. Set auto there and Claude Code go fall back to built-in default, skip any defaultMode for ~/.claude/settings.json, without printing any error anywhere. For version or plan wey built-in default na Manual, session go quietly start for default mode. Move the line go ~/.claude/settings.json. Run /permissions to list every active rule beside the file wey e come from.
Switch wey fit turn mode off
Administrators get two kill switches, and both dey use the string "disable" instead of boolean.
{
"permissions": {
"disableAutoMode": "disable",
"disableBypassPermissionsMode": "disable"
}
}Documentation specify say both settings accept the string "disable" for any settings file. Dem dey most useful for managed settings wey users no fit override.
disableAutoMode remove auto from Shift+Tab cycle, and session wey start with --permission-mode auto go start for Manual mode instead. disableBypassPermissionsMode do the same thing for bypass mode, and e work from any scope. So you fit set am for your own ~/.claude/settings.json to lock yourself out of mode wey you no wan reach by 2am for live server. For machine wey other people dey use, put both for /etc/claude-code/managed-settings.json instead, because user settings file belong to the user, but managed one no belong to user.
Wetin make auto mode for VPS need isolation boundary
The classifier dey review one action at a time. E no contain wetin approved action go do afterward. The documentation draw the line clearly:
The classifier na per-action control, no be isolation boundary, so isolation boundary still add defense in depth for unattended runs, and e no be required the way e be for --dangerously-skip-permissions.
So, for remote box, the correct combination na auto mode plus environment wey you dey ready to lose, no be bypassPermissions plus hope. Documentation only recommend bypass mode for isolated environments: containers, virtual machines, or dev containers wey no get internet access, where Claude Code no fit damage your host system. VPS wey dey run your database and reverse proxy no be any of these.
Three things carry most of the protection for server. Run Claude Code as normal user; never use root. Give that user working directory and nothing else worth reading. Rebuild the box instead of repairing am. Na this be the reason for disposable VM wey you throw away after each job. The hardening details behind these steps, from user creation to firewall rules, dey covered for the full safety pass for running Claude Code on a VPS, so we no repeat dem here.
Claude Code enforce the root rule by itself. For Linux and macOS, e refuse to start in bypass mode under sudo or as root:
--dangerously-skip-permissions cannot be used with root/sudo privileges for security reasonsThat check no dey run inside recognized sandbox. Na why the documented answer for autonomous container work be dev container wey dey run Claude Code as non-root user. If you dey control the agent from phone or laptop through SSH, the same reasoning apply to long-running Claude Code session wey you keep open inside tmux: nobody dey watch the prompt while the session dey run. Nothing fit start until the login itself work. So, if the box reject you with Permission denied (publickey), the verbose SSH output go show you which of the five faults you actually get.
Set up the Bash sandbox for one Ubuntu VPS
The built-in sandbox dey restrict filesystem and network access for sandboxed Bash commands wey Claude dey run, and operating system dey enforce the same restriction for child processes too. Excluded commands and approved unsandboxed retries dey run outside am. For Linux, e need two packages.
sudo apt-get install bubblewrap socatStart Claude Code and run /sandbox. The panel go open with Mode, Overrides and Config tabs. If any required package no dey, na Dependencies tab only go show, and e go list wetin you need install. Dependency check dey run when startup happen, so restart Claude Code after you install the packages. If you no restart am, the panel go still report say dem no dey.
For Ubuntu 24.04 and later, default AppArmor policy dey stop bubblewrap from creating the user namespaces wey e need, so sandbox no go start. Check whether this one affect your box:
sysctl kernel.apparmor_restrict_unprivileged_usernsA 0, or error wey talk say the key no exist, mean say you no need do anything. A 1 mean say bwrap need im own profile:
sudo tee /etc/apparmor.d/bwrap > /dev/null <<'EOF'
abi <abi/4.0>,
include <tunables/global>
profile bwrap /usr/bin/bwrap flags=(unconfined) {
userns,
include if exists <local/bwrap>
}
EOF
sudo systemctl reload apparmorThe profile apply to bwrap itself, no be to the commands wey e run inside sandbox. Then narrow the boundary for settings:
{
"sandbox": {
"enabled": true,
"filesystem": {
"denyRead": ["~/"],
"allowRead": ["."]
},
"network": {
"allowedDomains": ["github.com", "*.npmjs.org"]
}
}
}Put that block for the project's .claude/settings.json, because . dey resolve to project root only when you use project settings. If you put the same lines for ~/.claude/settings.json, . go resolve to ~/.claude instead. That one mean say denyRead rule go keep your project files blocked, and every command go fail to read the code wey e suppose edit.
Know wetin this no cover. Sandbox dey constrain Bash and im child processes. Built-in file tools dey run inside Claude Code process, while MCP (model context protocol) servers and hooks na separate processes wey dey run without restriction for host. To put all of dem behind one boundary, run the whole Claude Code process inside container, virtual machine, or the @anthropic-ai/sandbox-runtime package. As at the time of writing, that package na beta research preview.
Which mode each setup deserve?
A solo repo for your own laptop
Use auto, with ask rules for the actions wey you want to monitor. You dey for the keyboard, the classifier fallback fit reach you, and the damage dey limited to one machine wey you control. Na this case the 14 August default target.
A shared VPS
Use auto for each user, set am inside each user's own ~/.claude/settings.json, for a box where the account wey dey run Claude Code no be root and no fit read other users' work. Deploy disableBypassPermissionsMode as "disable" inside /etc/claude-code/managed-settings.json, together with deny rules wey protect shared paths. Shared box na the clearest case where bypassPermissions wrong, because the isolation boundary wey that mode assume no dey exist: the other tenants dey inside am.
CI and unattended sessions
Use dontAsk with explicit allow list of the commands wey the job need. Auto-deny na the correct failure mode when nobody go ever see prompt. Auto mode dey run non-interactively too, but -p run wey hit repeated classifier blocks no dey stop: e skip the blocked actions and Claude continue to work, so the job fit finish with some steps never done. Keep bypassPermissions for container or virtual machine wey you rebuild from image, and no use am for any host wey also dey run something wey matter to you.
FAQ
Auto mode go become default for Claude Code when?
E become default on 14 August 2026 for new sessions on Pro, Max, and Team plans. As of 27 October 2026, the documentation talk say Claude Code v2.1.283 or later get am as the built-in starting mode for interactive terminal and VS Code sessions on every plan and provider. The exception na organizations wey apply HIPAA configuration; dem start for Manual mode with v2.1.285 or later. Non-interactive claude -p runs start for default unless you set a mode. Before v2.1.302, runs for third-party provider or with telemetry off fit start for auto. From v2.1.302, dem start for default too. Any default wey you set yourself go remain unless you accept the one-time switch prompt. The default wey your organization manage no change. Check the mode wey session dey use by looking the status bar. For auto mode, e go show ⏵⏵ auto mode on.
I suppose use auto mode or bypassPermissions for VPS?
Use auto mode with isolation boundary. The classifier go review actions before dem run. Reads and working-directory edits outside protected paths go skip review. But the documentation clear say this na per-action control, no be isolation boundary. So unattended run still need container, virtual machine, or server wey you ready rebuild. bypassPermissions skips all checks, and the documentation say make you use am only for isolated environments. Claude Code no go start for that mode as root on Linux. E go print --dangerously-skip-permissions cannot be used with root/sudo privileges for security reasons.
How I fit stop anybody for my server from using auto mode or bypass mode?
Set permissions.disableAutoMode and permissions.disableBypassPermissionsMode to string "disable" inside /etc/claude-code/managed-settings.json. Managed settings dey above every other scope. So no user settings file or command line flag fit override dem. disableAutoMode removes auto from Shift+Tab cycle. A session wey start with --permission-mode auto go start for Manual mode instead. disableBypassPermissionsMode dey work from any scope too, so one user fit set am inside their own ~/.claude/settings.json.
Why my defaultMode: "auto" setting dey ignored?
Because e dey inside wrong file. defaultMode: "auto" no dey take effect when e come from .claude/settings.json or .claude/settings.local.json. This stop repository from granting itself auto mode by shipping settings file. Claude Code go use its built-in default instead. E go skip any defaultMode inside your user settings and print no error. If the built-in default na Manual, the session go start for default mode. Move the setting go ~/.claude/settings.json. Then run /permissions to confirm the file wey each active rule come from. If auto mode still no dey available, check the model requirement. Older models like Sonnet 4.5 no get support for any provider.