| name | audit-claude-config |
| description | Use when needing a full inventory of Claude Code plugins, MCP servers, skills, permissions, and integrations across all configuration levels. Triggers on "what do I have installed", "audit my setup", "what MCP servers do I have", "what plugins", "show my config", "what's configured where", "claude setup overview" |
Audit Claude Configuration
Scan all Claude configuration levels and produce a structured inventory of plugins, MCP servers, skills, permissions, and integrations.
When to Use
- User wants to understand what's installed/configured across their Claude setup
- Auditing for security (hardcoded credentials, overly broad permissions)
- Cleaning up stale or duplicate configurations
- Onboarding to an existing machine's Claude setup
Configuration Levels
Scan these locations in order. Each level can override or extend the ones above it.
Level 1: User-Global
| Location | What to Check |
|---|
~/.claude/settings.json | Enabled plugins, permissions, hooks, voice, status line |
~/.claude/CLAUDE.md | Global instructions |
~/.claude/plugins/installed_plugins.json | Full plugin inventory with versions |
~/.claude/plugins/blocklist.json | Blocked plugins |
~/.claude/skills/ | Personal skills (list dirs) |
~/.claude/keybindings.json | Custom key bindings |
~/.claude/teams/ | Team configurations (if present) |
~/.claude/ide/ | IDE integration configs (VS Code, JetBrains) |
~/.claude/mcp-needs-auth-cache.json | MCP servers pending authentication |
~/.claude/security_warnings_state_*.json | Suppressed security warnings |
Level 2: Claude Desktop App
| Location | What to Check |
|---|
~/Library/Application Support/Claude/claude_desktop_config.json (macOS) | MCP servers, trusted folders, cowork settings, preferences |
%APPDATA%/Claude/claude_desktop_config.json (Windows) | Same as above |
~/Library/Application Support/Claude/Claude Extensions/ (macOS) | Installed desktop extensions |
~/Library/Application Support/Claude/extensions-installations.json | Extension install tracking |
~/Library/Application Support/Claude/extensions-blocklist.json | Blocked extensions |
~/Documents/Claude/ | Cowork projects and scheduled tasks |
Level 3: Project-Level
Search the user's project directories (ask if not obvious):
| Pattern | What to Check |
|---|
<project>/.claude/settings.local.json | Project permissions, enabled MCP servers, experimental flags |
<project>/.claude/settings.json | Shared project settings (committed to git) |
<project>/.mcp.json | Project MCP server definitions (servers, credentials, args) |
<project>/CLAUDE.md | Project-specific instructions |
<project>/.claude/skills/ | Project-specific skills |
<project>/.claude/commands/ | Custom slash commands |
Level 4: User-Project Overrides
| Location | What to Check |
|---|
~/.claude/projects/<hashed-path>/ | Per-project settings stored at user level (not in repo) |
Execution Strategy
Use 2-3 parallel tool calls or subagents for speed:
- Batch 1: User-global + Desktop app (Levels 1-2)
- Batch 2: All project-level configs (Level 3) — search per known directory rather than a single deep glob (deep
** globs can time out on large trees)
- Batch 3 (if needed): Deep-read large configs found in Batch 2
Inferring project locations: Check the working directory, ~/Developer, ~/Projects, ~/repos, ~/code. If none contain projects, ask the user. In a subagent context where you cannot ask, search common locations and note which were checked.
Output Format
Structure the report as:
1. Summary Table (one row per config level)
| Level | Plugins | MCP Servers | Skills | Key Flags |
2. Plugin Inventory
List all installed plugins with source, version, and whether actively used or potentially stale.
3. MCP Server Map
For each unique MCP server found:
- Name and type (Docker, npx, remote, etc.)
- Which levels/projects it appears in (flag duplicates)
- Credentials present (YES/NO - never print actual secrets)
- Auth status (authenticated, pending, unknown)
Include a separate row for Desktop Extensions (these are distinct from project-level MCP servers).
4. Per-Project Breakdown
Table with columns: Project | MCP Servers | Permissions Count | Notable Config
Group by project directory (e.g., work vs. personal) when the user has distinct top-level directories.
5. Flags and Recommendations
Group into:
Security: Hardcoded credentials, overly broad permissions, suppressed warnings
Duplication: MCP servers or configs repeated across projects that could be global
Stale: Empty configs, old worktrees, unused plugins, pending auth
Cost: Docker-based MCP servers, plugins that may not be needed
Common MCP Server Types
These appear frequently and are candidates for global config:
- Context7 (docs lookup) - almost always duplicated per-project
- Browser-Tools / Puppeteer - browser automation
- Atlassian (Docker) - Jira/Confluence
- Supabase - database
- Firebase - backend
- Filesystem - local file access
- AWS MCP servers - cloud operations
Security Checks
Flag but never print actual values for:
- API tokens/keys in
.mcp.json files
- Service role keys in settings
- Database connection strings with credentials
- Database passwords in MCP configs
- OAuth tokens
Report as: "Atlassian API token found in plaintext in 7 project .mcp.json files"
When reading configs that may contain secrets: Extract only key names and structure. If a config file has been read and contains plaintext secrets, do not reproduce them in the output. Summarize as "credentials present: YES" with the credential type.
