MCP Config Security Checkup — 14 static checks

Before you wire an MCP (Model Context Protocol) server config into your agent, run it through a static security checkup first. This skill runs fully on your machine by default — the config never leaves the device and no network connection is ever made — and executes 14 automated static checks, reporting findings as CRIT/HIGH/MED/LOW with concrete fix suggestions.

中文：在把 MCP 服务器配置挂到 Agent 前先做静态安全体检，默认全程本机运行、 零网络、配置不外传，输出 CRIT/HIGH/MED/LOW 分级发现与修复建议。

What it checks

Pure static analysis of an MCP config document (usually a JSON containing mcpServers; a single remote endpoint URL is also accepted). It never connects to any address found in the config. The 14 checks cover:

• CRIT: cloud credential/secret signature strings hardcoded in the config; filesystem servers that are writable or mount root/home; exposure of arbitrary shell/command execution.
• HIGH: generic code/expression evaluation tools; fetch/network servers missing host allowlists (SSRF surface); generic secret/token strings in environment variables; database servers not declared read-only.
• MED: plaintext-HTTP remote endpoints; command strings with injection patterns ($(...), backticks, pipe chaining); configs clustering exec/file/net dangerous capabilities; endpoints pointing at loopback, intranet, or cloud metadata addresses.
• LOW: npx/uvx remote package launchers with unpinned versions (supply-chain surface).

Outputs risk_level (CRIT/HIGH/MED/LOW/PASS), per-level counts, rule IDs, server names, redacted evidence, and remediation advice.

What it does NOT do

• Zero network by default: every check runs in a local process; the config document is never uploaded or sent anywhere, and no URL/endpoint in the config is ever contacted.
• No dynamic probing: even in --endpoint-url mode the URL is merely wrapped into a single-server config for static text checks — no connection is made. Live endpoint behavior, tool implementation code, permission boundaries, and data flows require a manual deep audit.
• It does not replace manual audit or penetration testing; the 14 automated checks are a first-pass screening layer, not a complete security assessment.
• It claims no endorsement by OWASP, IETF, or any standards body; standard names are referenced only to explain the check basis.

Usage

# Check up an MCP config file (a document containing mcpServers)
python3 scripts/mcp_checkup.py -c examples/mcp_dangerous.json

# Static wrap check for a single remote endpoint URL (no connection made)
python3 scripts/mcp_checkup.py -u https://example.com/mcp

# Machine-readable JSON
python3 scripts/mcp_checkup.py -c mcp.json --json

# Read from stdin
cat mcp.json | python3 scripts/mcp_checkup.py

Exit codes: 0 = check ran to completion (any risk level counts as a finding, including CRIT); 2 = input error; 3 = online mode returned HTTP 402.

Arguments:

Optional online mode: by default every check runs locally with zero network and zero data collection. Only with an explicit --online flag is the config submitted to the Correctover CCS hosted service. That hosted service is currently metered via Alipay A2M (Chinese payment rail); for overseas users the local mode is full-featured and free with no call limits, and an x402/USDC endpoint is on the roadmap. The local path needs no credentials.

Example

examples/mcp_dangerous.json is a deliberately dangerous config (root filesystem mount, hardcoded cloud keys, shell executor, plaintext HTTP):

  risk_level: CRIT
  Findings: CRIT=3  HIGH=3  MED=4  LOW=2
  Automated checks: 14
  [CRIT] Cloud credential/secret hardcoded in config          (MCP-001)
  [CRIT] Filesystem server writable or mounts root/home       (MCP-002)
  [CRIT] Arbitrary shell / command execution exposed          (MCP-003)
  [HIGH] Network/fetch server missing host allowlist (SSRF)   (MCP-005)
  ...

Secret values in evidence are redacted (e.g. AKIAIO…(redacted)); full sensitive values are never echoed.

Requirements

Python 3.7+, standard library only (json/re/time). No third-party packages. Default mode makes no network requests.

Privacy & security

• Pure static analysis, local by default: the config document never leaves the machine; no URL in the config is contacted.
• The script collects and uploads nothing; it is stateless and executes no commands against the target.
• Only explicit --online submits the config to the Correctover hosted service (self-managed API key). The server logs no request bodies — only an anonymous request ID, endpoint name, success/failure, and latency; request bodies are capped at 2 MiB.

About

By Correctover (Guigui Wang) — Correctover = AI Reliability. We build agent runtime verification: CCS (Correctover Conformance Shape) receipts, Ed25519 tool-call verification, and AI-agent security tooling. Homepage: https://correctover.com/?utm_source=clawhub&utm_medium=skill&utm_campaign=ccs_checkup

Disclaimer

This report is auto-generated and applies only to the submitted input. Automated checks are an aid only and constitute no security guarantee, compliance conclusion, audit opinion, or legal opinion. No endorsement by OWASP, IETF, or any standards body is claimed. The live behavior of remote endpoints, tool implementation code, permission boundaries, and data flows are outside the scope of automated checks and require manual review. Any security decision must ultimately rely on qualified human review.

License

MIT-0 for this ClawHub distribution. Anyone may use, modify, and redistribute it, including commercially, with no attribution required.