Off Switches

Off switches

This page lists what the settings snippet turns off in a vanilla Claude Code install, what it never sets, and what has no gate. The research behind it is the off-switch survey of Claude Code 2.1.293 and 2.1.294. Each key and env entry has its reason on Settings snippet keys. dotclaude writes none of these keys. You copy the ones that you want.

Forced off, no quality cost

A scope of User means that Claude Code reads the key from user settings and local settings, and ignores it in project settings. Put such a key in user settings.

Part	Key or entry	Scope	Reason
Built-in plugins	enabledPlugins with cc-plugin-<name>@builtin set to false	Any	No use in dotclaude, or a panel or tip that adds noise.
Spinner tips	spinnerTipsEnabled	Any	No use.
Survey	feedbackSurveyRate	Any	No use.
Feedback tool	feedbackDrafts	Any	Removes the SendFeedback tool.
Fast mode	fastMode	Any	Costs more.
Agent view	disableAgentView	Any	Removes a daemon.
Remote Control	disableRemoteControl	Any	Removes a remote surface.
claude.ai skills	syncClaudeAiSkills	User	No anthropic-skills:* skill in the listing.
claude.ai plugins	syncClaudeAiPlugins	User	No synced plugin.
claude.ai connectors	disableClaudeAiConnectors	Any	No claude.ai MCP server.
computer-use server	permissions.deny entry mcp__computer-use	Any	See Computer use.
Chrome prompt	CLAUDE_CODE_DISABLE_CFC_PROMPT	Any	Removes the Chrome part of the prompt.
Cron tools	CLAUDE_CODE_DISABLE_CRON	Any	Removes the tools and /loop.
Advisor tool	CLAUDE_CODE_DISABLE_ADVISOR_TOOL	Any	Removes the tool and its text.
Terminal title	CLAUDE_CODE_DISABLE_TERMINAL_TITLE	Any	Removes a small-model request.
Attribution header	CLAUDE_CODE_ATTRIBUTION_HEADER set to 0	Any	Removes a block at the start of the prompt.
IDE auto-connect	CLAUDE_CODE_AUTO_CONNECT_IDE set to false	Any	The key is in the global config, so the snippet uses the variable.
IDE extension install	CLAUDE_CODE_IDE_SKIP_AUTO_INSTALL	Any	The global config holds the key.
Task list reminder	CLAUDE_CODE_TODO_REMINDER_MODE set to off	Any	Removes a recurring reminder.
Built-in agents	permissions.deny entries Agent(claude-code-guide) and Agent(statusline-setup)	Any	Removes two agent lines from each request.
Plan mode tools	permissions.deny entries EnterPlanMode and ExitPlanMode	Any	Removes two tool lines. The eval of 2026-10-08 showed no loss, see Not in the snippet.
Task tools	permissions.deny entries TaskCreate, TaskGet, TaskList, TaskStop, and TaskUpdate	Any	Saves about 30 tokens, and the model keeps no task list. The eval showed no loss.
Git instructions	includeGitInstructions set to false	Any	About 6.2k characters of git text and the git status snapshot leave the prompt. The eval showed no loss, and the Bash tool text of features/surface/data.mjs carries the attribution line.

The built-in plugins that the snippet turns off are cc-plugin-agents-md, cc-plugin-claude-test, cc-plugin-diff, cc-plugin-mermaid, cc-plugin-mods-guide, cc-plugin-responsive-mode, cc-plugin-telemetry, cc-plugin-tips, and cc-plugin-you-should-know. The snippet leaves cc-plugin-plugin-authoring on, because it teaches the plugin format. It leaves cc-plugin-sec-default alone, because only a managed setting enables it. A new built-in plugin stays as Claude Code sets it.

The skillOverrides entry plugin-authoring is name-only. Claude Code skips skillOverrides for a skill from a plugin (task 2.12). Whether it skips a skill of a built-in plugin is not verified. Check it with /context in just sandbox.

Built-in agents

The five dotclaude agents take the work of the built-in agents (Agents). The snippet turns off these built-in agents:

• Explore and Plan, with CLAUDE_CODE_DISABLE_EXPLORE_PLAN_AGENTS=1.
• general-purpose, claude-code-guide, and statusline-setup, with one permissions.deny entry Agent(<name>) each.

Each listed agent adds a line to each request, and Explore runs on the main model. With the snippet, the agent listing holds only the five dotclaude agents.

Computer use

The computer-use MCP server is opt-in. Claude Code starts it only when the project config key enabledMcpServers holds computer-use. In 7 sandbox runs on 2.1.293 the server was never in a request, so the default needs no switch. The mcp__computer-use deny rule is a guard for the case that you enable it. The deny rule on a running server is not tested, because the server needs the cu-desktop binary.

Not in the snippet

These items might cost quality, so they are not in the snippet. The harness that tested them is gone, and the format of claude plugin eval has no gate for them (design D9, Evals). So no gate exists for these items, and they stay off. Set one by hand in your own settings at your own risk.

Item	What it sets	Saving
Telemetry	DISABLE_TELEMETRY set to 1, which also stops the GrowthBook fetch	No flag fetch, so flags use the code defaults.
Non-essential traffic	CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC set to 1	No auto-update, error report, release notes, or PR badge check.
Subagent cache of 1 hour	subagentPromptCacheTtl set to 1h	See the reason below.

Reason for the subagent cache: the measure of 2026-10-05 (981 subagent runs in 7 days, see Design) found that 1 hour costs $171 more than 5 minutes. Of 21,807 gaps between two requests, 41 were 5 to 60 minutes long.

Eval of 2026-10-08

The eval ran 12 cases with 3 runs per arm, with the harness of 0.27.0 and the first 0.28.0 work. The rule was to drop an item when the resolve rate, the over-claim rate, or the policy-violation rate got worse beyond the noise bound. When another measure got worse with a shown interval, the arm ran again, and the item was dropped only when the loss showed again. This page gives the figures, because the reports are not in the repository.

Item	Result	Decision
plan-mode	First run: multiSentenceShare 0.6 against 0.718, shown. Second run: 0.694 against 0.752, not shown. Cost was lower with the item (0.002, interval 0 to 0.003, shown).	In the snippet.
task-tools	No measure was worse.	In the snippet.
git-instructions	No measure was worse. A sandbox commit kept the Co-Authored-By trailer.	In the snippet.
telemetry	Second run: turns 4.75 against 5.5, seconds 12.113 against 15.432, and outputTokens 655.861 against 799.917. All three were shown worse for the item.	Not in the snippet.
nonessential-traffic	Second run: turns 4.75 against 5.389 and outputTokens 655.861 against 762.472. Both were shown worse for the item.	Not in the snippet.
subagent-cache-1h	Cost for each resolved task 0.028 against 0.027. The mean difference was 0.002, interval -0.008 to 0.007. No saving was shown.	Not in the snippet.

The likely cause of the telemetry loss is that DISABLE_TELEMETRY stops the GrowthBook flag fetch. This cause is not verified.

Opt-in

dotclaude 0.28.0 has no opt-in item. Auto-fix at Stop and its DOTCLAUDE_AUTOFIX variable are removed, so no hook runs a project formatter outside the sandbox.

Never set

The snippet never sets a key that stops hooks or plugins, removes project instructions or attachments, or forces a model or effort. The list is a constant in tests/dotclaude/settings-pin.test.mjs, because lib/ ships to users and only the test reads it. The test fails when the plugin settings.json or any JSON block of the snippet page sets any entry.

Entry	Reason
disableAllHooks	Stops all hooks.
allowManagedHooksOnly	Stops the hooks of plugins.
strictPluginOnlyCustomization	Stops customization by the user.
disableSideloadFlags	Stops the flags that load a plugin folder.
alwaysThinkingEnabled	Forces thinking on every model.
allowAllClaudeAiMcps	Turns on every claude.ai connector.
claudeMd	Replaces the project instructions.
CLAUDE_CODE_SIMPLE	Turns off hooks, plugins, skills, and memory.
CLAUDE_CODE_SAFE_MODE	Turns off hooks, plugins, skills, and memory.
CLAUDE_CODE_ENABLE_TASKS set to 0	Removes the task tools that the model uses.
CLAUDE_CODE_DISABLE_CLAUDE_MDS	Removes the project instructions.
CLAUDE_CODE_DISABLE_ATTACHMENTS	Removes the attachments.
CLAUDE_CODE_EFFORT_LEVEL	Forces one effort on every model.
CLAUDE_CODE_SUBAGENT_MODEL_FORCE	Forces the model of a subagent.
MAX_THINKING_TOKENS	Forces a thinking budget.
CLAUDE_CODE_REMOVE_PROMPT_STRINGS	Removes text of the system prompt.
CLAUDE_CODE_DISABLE_WEB_FETCH	Removes the web fetch tool.
CLAUDE_CODE_BASH_EDIT_DIFF set to 0	Removes the edit diff of Bash.
CLAUDE_CODE_SKIP_PLUGIN_MCP_SERVERS	Drops the MCP servers of plugins.
CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS	Removes beta features.
DISABLE_INTERLEAVED_THINKING	Removes thinking between tool calls.

Managed settings

dotclaude uses no managed settings. It writes no managed-settings.json and no managed-settings.d/ drop-in, runs no sudo, and writes nothing in the Claude Code install folder. The 0.27.0 template templates/managed-settings.json is gone. dotclaude no longer reads, moves, or removes a managed file that an earlier version wrote. If you have one, review it yourself.

Deny-override trade-off

With no managed settings, the guard cc-plugin-sec-default is not seated, and dotclaude runs at tier user. An installed mod with a tool.check hook can then turn a content deny rule into an allow. The probe turned WebFetch(domain:example.com) from deny into allow this way. A path deny such as Read(./secret.txt) is enforced before tool.check, and a whole-tool deny removes the tool, so no mod can override those two. You accept this trade-off when you install a mod. The guard of dotclaude never weakens a deny.

Organization case

On a machine with managed settings, or with a Team or Enterprise login, the guard is seated outermost. It skips the prompt.section handlers of tier user, so the prompt text of dotclaude does not reach the request. The SessionStart drift check reports that the guard is seated and which parts it skips. dotclaude changes nothing else.