Settings Management

Manage configuration for coding agents.

IMPORTANT: After modifying settings, always inform the user that they need to restart the agent for changes to take effect. Most settings are only loaded at startup.

Settings File Locations

Precedence (highest to lowest): Managed → Command line → Local → Project → User

Quick Actions

View Current Settings

cat ~/.claude/settings.json 2>/dev/null || echo "No user settings"
cat .claude/settings.json 2>/dev/null || echo "No project settings"
cat .claude/settings.local.json 2>/dev/null || echo "No local settings"

Create/Edit Settings

Use the Edit or Write tool to modify settings files. Always read existing content first to merge changes.

Common Configuration Tasks

Set Default Model

{
  "model": "claude-opus-4-7",
  "effort": "high"
}

Effort levels: low, medium, high, xhigh (Opus 4.7 only), max. As of 2026, default effort is high for API-key, Bedrock/Vertex/Foundry, Team, and Enterprise users.

Configure Permissions

{
  "permissions": {
    "allow": ["Bash(npm run:*)", "Bash(git:*)"],
    "deny": ["Read(.env)", "Read(.env.*)", "WebFetch"],
    "defaultMode": "acceptEdits"
  }
}

defaultMode accepts: default, acceptEdits, plan, auto, dontAsk, bypassPermissions. The new auto mode (March 2026) uses an LLM-based classifier and triggers PermissionDenied hooks on rejection.

Add Environment Variables

{
  "env": {
    "MY_VAR": "value",
    "CLAUDE_CODE_ENABLE_TELEMETRY": "1"
  }
}

Enable Extended Thinking

{
  "alwaysThinkingEnabled": true
}

Configure Attribution

{
  "attribution": {
    "commit": "Generated with AI\n\nCo-Authored-By: AI <ai@example.com>",
    "pr": ""
  }
}

Configure Sandbox

{
  "sandbox": {
    "enabled": true,
    "autoAllowBashIfSandboxed": true,
    "excludedCommands": ["docker", "git"]
  }
}

Configure Hooks

{
  "hooks": {
    "PreToolUse": {
      "Bash": "echo 'Running command...'"
    }
  }
}

Scope Selection Guide

• User settings (~/.claude/settings.json): Personal preferences across all projects
• Project settings (.claude/settings.json): Team-shared settings, commit to git
• Local settings (.claude/settings.local.json): Personal project overrides, not committed

Workflow

1. Determine scope: Ask user which scope (user/project/local) if not specified
2. Read existing settings: Always read current file before modifying
3. Merge changes: Preserve existing settings, only modify requested keys
4. Validate JSON: Ensure valid JSON before writing
5. Confirm changes: Show user the final settings
6. Remind to restart: Tell user to restart Claude Code for changes to take effect

Codex CLI Settings

Codex uses TOML format in ~/.codex/config.toml (user) and .codex/config.toml (project, trusted projects only).

model = "gpt-5.5"                  # As of April 2026; gpt-5.4 is a valid fallback
approval_policy = "on-request"     # untrusted | on-request | never (or { granular = { ... } })
                                    # NOTE: "on-failure" is DEPRECATED — migrate to on-request or never
sandbox_mode = "workspace-write"   # read-only | workspace-write | danger-full-access

[features]
codex_hooks = true                 # Lifecycle hooks (stable in v0.124, April 2026)
multi_agent = true                 # Subagent orchestration (GA March 2026)

Key differences from Claude Code:

• TOML format instead of JSON; project config requires explicit trust
• Starlark rules for command policies in .codex/rules/
• Named profiles ([profiles.NAME]) for different workflows
• Feature flags system (codex features list, codex --enable feature)
• Lifecycle hooks live inline as [[hooks.PreToolUse]] (etc.) blocks in config.toml
• [agents] block for subagent orchestration (max_threads, max_depth)
• [[skills.config]] for per-skill enable/disable overrides
• Custom model providers via [model_providers.NAME], including amazon-bedrock since v0.123

See references/codex-settings.md for the full Codex config reference (covers approval/sandbox/profiles/features/agents/skills/hooks/rules/providers/admin enforcement).

OpenCode Settings

OpenCode (anomalyco/opencode v1.14.x) uses JSON/JSONC in ~/.config/opencode/opencode.json (user) and opencode.json (project root, or under .opencode/).

{
  "$schema": "https://opencode.ai/config.json",
  "model": "anthropic/claude-sonnet-4-5",
  "permission": {
    "edit": "ask",
    "bash": { "*": "ask", "git status *": "allow" }
  },
  "instructions": ["AGENTS.md", "docs/style.md"]
}

Key differences from Claude Code:

• Schema-validated JSON/JSONC, not plain JSON
• Configs are deep-merged (later wins; arrays like instructions are concatenated, not replaced)
• Permissions are an object of allow/ask/deny per tool with glob patterns, not separate allow/deny/ask arrays
• Theme/keybinds live in a separate tui.json file
• Multi-provider via model: "<provider>/<model-id>" and a top-level provider block
• Variable substitution: {env:VAR} and {file:path}

See references/opencode-settings.md for full OpenCode config reference.

Devin CLI / Desktop Settings

Devin uses JSON in ~/.config/devin/config.json (user; %APPDATA%\devin\config.json on Windows), .devin/config.json (project), and .devin/config.local.json (project-local, gitignored). MCP servers live in dedicated mcp_config.json files at the same levels.

{
  "agent": { "model": "swe-2-high" },
  "permissions": {
    "allow": ["Read(**)", "Exec(git)"],
    "ask": ["Write(**/.env*)"]
  },
  "read_config_from": { "claude": true, "cursor": true, "windsurf": true }
}

Key differences:

• Project configs accept only permissions, read_config_from, and hooks
• read_config_from imports rules/hooks/subagents from .claude/, .cursor/, .windsurf/ by default
• Skills: .devin/skills/ (project), ~/.config/devin/skills/ (user)
• Plugins: .devin-plugin/plugin.json manifest; devin plugins CLI

See references/devin-settings.md for the full Devin config reference.

Reference

• Claude Code settings: references/claude-settings.md
• Codex CLI settings: references/codex-settings.md
• OpenCode settings: references/opencode-settings.md
• Devin CLI/Desktop settings: references/devin-settings.md