Mcp Audit

MCP Audit

Inventory and audit every Model Context Protocol server the user has configured — both user-scope and project-scope — read through the Agent Monitor dashboard at http://localhost:4820.

Input

The user provides: $ARGUMENTS

This may be:

• empty — audit all MCP servers (default).
• a server name fragment — focus on matching servers.
• "stdio" / "http" — restrict to one transport kind.

Data Sources

| Endpoint | Returns | |----------|---------| | GET /api/cc-config/mcp | { user:[…], projectScoped:[…] }. Each server: { name, source, kind } where kind is stdio (with command, args, envNames), http (with url, headers), or unknown. source names the file the definition came from (e.g. ~/.claude.json (top-level), ~/.claude.json (projects[]), ~/.claude/settings.json) |

Report Sections

1. Server inventory

List every server from user and projectScoped. For each show name, source, kind, and the transport detail:

• stdio — the command, its args, and the envNames (names only — values

are not exposed by the API).

• http — the url and the headers key names (values not exposed).
• unknown — a definition the server could not classify; flag it for review.

2. Scope split & duplication

Separate user-scope from project-scope servers. Flag any name that appears in both scopes (project may shadow user) and any duplicate definitions across source files.

3. Hygiene flags

• Unknown transport — servers with kind: "unknown" (malformed or

unsupported definition).

• Env reliance — stdio servers with many envNames; note they depend on

environment variables being present at launch.

• Remote endpoints — http servers; surface the url host so the user can

confirm they trust the remote.

Output

• Section 1 as a table (Scope | Name | Kind | Transport detail | Source).
• Env names and header names listed by name only — never invent or print values

(the API does not expose them).

• Cite only fields the API returned — never fabricate servers, commands, or

hosts.

• Note: MCP servers are read-only via the Config Explorer (they are written

concurrently by the running CLI); edit their definitions in the source file named by source.

• If the dashboard is unreachable at http://localhost:4820, say so and tell

the user to start it with npm start from the repo root.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: hoangsonww
• Source: hoangsonww/Claude-Code-Agent-Monitor
• License: MIT
• Homepage: https://hoangsonww.github.io/Claude-Code-Agent-Monitor/

Install and usage instructions live in the source repository linked above.