settings-reference
Claude Code settings.json reference for permissions, sandbox, plugins, skill visibility, and configuration options. Claude loads automatically when asked about settings, permissions, configuration, or hooks.
- License 许可
- GPL-3.0
- 最近更新
- 2026-07-22
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- settings-reference
- description
- Claude Code settings.json reference for permissions, sandbox, plugins, skill visibility, and configuration options. Claude loads automatically when asked about settings, permissions, configuration, or hooks.
- when_to_use
- settings, permissions, deny, allow, sandbox, hooks, plugin, skillOverrides, env, defaultMode, availableModels, companyAnnouncements, footerLinksRegexes, configuration, managed settings, disableBundledSkills
- user-invocable
- false
03 SKILL.md 正文
Claude Code Settings Reference
Settings precedence (highest → lowest)
- Managed (MDM / server / file) — cannot be overridden by any lower scope
- CLI flags /
--settings <file> - Local project
.claude/settings.local.json - Shared project
.claude/settings.json - User
~/.claude/settings.json
Array settings merge across scopes (exception: fallbackModel and availableModels do not merge).
Key settings
| Key | Effect |
|---|---|
permissions.allow/ask/deny |
Tool permission rules |
permissions.defaultMode |
default / acceptEdits / plan / auto / bypassPermissions |
env |
Environment variables injected every session |
model |
Default model override |
effortLevel |
Persist effort: low / medium / high / xhigh |
availableModels |
Restrict which models users can select |
skillOverrides |
Per-skill: on / name-only / user-invocable-only / off |
disableBundledSkills |
Remove bundled skills (code-review, debug, loop, etc.) |
disableSkillShellExecution |
Block !cmd`` in skills from user/project sources |
autoCompactEnabled |
Auto-compact when context approaches limit |
sandbox.enabled |
Isolate bash from filesystem and network |
sandbox.network.allowedDomains |
Network allowlist for sandboxed commands |
hooks |
Lifecycle automation (PreToolCall, PostToolCall, Stop, etc.) |
companyAnnouncements |
Startup messages shown to all users |
footerLinksRegexes |
Turn IDs in output into clickable footer badges |
strictPluginOnlyCustomization |
Lock skills/agents/hooks/mcp to plugin sources only |
agent |
Run main thread as a named subagent |
Permission rule syntax
Bash(npm run *) # prefix match on command
Read(./.env) # exact file path
WebFetch(domain:*.com) # domain wildcard
Skill(deploy *) # skill name prefix
mcp__github__get_* # MCP tool glob
Deny rules are checked first, then ask, then allow. First match wins.
Skill visibility (skillOverrides)
{
"skillOverrides": {
"legacy-context": "name-only",
"deploy": "off"
}
}
on → full description in context; name-only → name only (saves budget); user-invocable-only → hidden from Claude; off → hidden everywhere.
Sandbox example
{
"sandbox": {
"enabled": true,
"filesystem": { "allowWrite": ["/tmp/build", "~/.kube"] },
"network": { "allowedDomains": ["github.com", "*.npmjs.org"] }
}
}
04 同仓库更多 Skills
workspace-visualize
Generate an interactive collapsible HTML tree view of the codebase with file sizes, type breakdown bar chart, and color-coded file types. Opens in your default browser. Use when exploring a new repo, understanding project structure, or identifying large files before refactoring.
testing-strategy-review
暂无简介
technical-domain-routing
暂无简介
target-codebase-overview
暂无简介
skill-forge
暂无简介
session-logger
Append a timestamped note to this session's activity log. Use to record decisions, key findings, or milestones during long sessions so the log outlasts context compaction.