Skip to content

Claude Code Advanced Configuration

Sandbox — OS-Level Isolation

Sandbox restricts filesystem and network access for all bash commands:

{
  "sandbox": {
    "enabled": true,
    "autoAllowBashIfSandboxed": true,
    "excludedCommands": ["docker *"],
    "filesystem": {
      "allowWrite": ["/tmp/build", "~/.kube"],
      "denyRead": ["~/.aws/credentials"]
    },
    "network": {
      "allowedDomains": ["github.com", "*.npmjs.org"],
      "allowLocalBinding": true
    }
  }
}
Setting Purpose
autoAllowBashIfSandboxed Auto-approve bash when sandbox is active
excludedCommands Commands that bypass sandbox (e.g. docker)
filesystem.allowWrite Additional writable paths
filesystem.denyRead Block reads to sensitive paths
network.allowedDomains Outbound network whitelist

Plugins

Plugins bundle skills, hooks, subagents, and MCP servers into a single installable unit:

/plugin                           # browse marketplace
/plugin marketplace add acme-corp/plugins

Plugin Settings

{
  "enabledPlugins": {
    "formatter@acme-tools": true,
    "deployer@acme-tools": true
  },
  "extraKnownMarketplaces": {
    "acme-tools": {
      "source": {
        "source": "github",
        "repo": "acme-corp/claude-plugins"
      }
    }
  }
}

MCP Servers

Model Context Protocol servers extend Claude with external tools:

claude mcp add                    # connect a new MCP server
MCP Server What It Provides
Context7 Auto-fetches library documentation
Chrome DevTools Live browser state, DOM, console, network
PostgreSQL Database schema and query results
GitHub Issues, PRs, and repository context
Filesystem File access and search

MCP tools appear in hook matchers as mcp__<server>__<tool>: - mcp__memory__create_entities - mcp__github__search_repositories

Non-Interactive Mode

Run Claude without a session for CI, scripts, and automation:

claude -p "Explain what this project does"
claude -p "List all API endpoints" --output-format json
claude -p "Analyze this log" --output-format stream-json

Fan-Out Pattern

Distribute work across parallel invocations:

for file in $(cat files.txt); do
  claude -p "Migrate $file from React to Vue. Return OK or FAIL." \
    --allowedTools "Edit,Bash(git commit *)"
done

Auto Mode for Unattended Execution

claude --enable-auto-mode          # enable once
claude --permission-mode auto -p "fix all lint errors"

Auto Mode uses a background safety classifier that runs independently on every action:

Layer What It Does
Input Scans tool outputs for prompt injection before they enter agent context
Output Evaluates each action via transcript classifier — blocks mass deletion, data exfiltration, scope escalation

Availability depends on plan and workspace policy. It is designed to reduce approval fatigue while still blocking destructive operations.

Enterprise Managed Settings

Centralized control for organizations:

Delivery Location
Server-managed Anthropic admin console
macOS MDM com.anthropic.claudecode managed preferences
Windows GPO HKLM\SOFTWARE\Policies\ClaudeCode
File-based /etc/claude-code/managed-settings.json (Linux)

Managed settings cannot be overridden by user or project settings.

Key Enterprise Controls

{
  "disableBypassPermissionsMode": "disable",
  "disableAutoMode": "disable",
  "allowManagedHooksOnly": true,
  "allowManagedPermissionRulesOnly": true,
  "strictKnownMarketplaces": [
    { "source": "github", "repo": "acme-corp/approved-plugins" }
  ]
}

Useful Settings Reference

Setting Purpose Example
model Override default model "claude-sonnet-4-6"
language Response language "japanese"
attribution Customize git commit attribution {"commit": "AI-generated", "pr": ""}
env Environment variables for every session {"FOO": "bar"}
autoUpdatesChannel "stable" (week-old) or "latest" "stable"
includeGitInstructions Disable built-in git workflow prompt false
plansDirectory Where plan files are stored "./plans"

JSON Schema Validation

Add to settings.json for IDE autocomplete:

{
  "$schema": "https://json.schemastore.org/claude-code-settings.json"
}

References

See also