Part of the StudioMeyer MCP Stack — Built in Mallorca 🌴 · ⭐ if you use it

mcp-stdio-shellguard

with allowlist + sandbox + replay-detection, plus an AST audit CLI (mcp-shellguard-audit) that scans MCP server sources for unsanitized shell calls. Closes the Ox-Security MCP stdio-RCE class (200k vulnerable servers, May 2026 disclosure).

• MCP spec: 2025-06-18
• SDK: @modelcontextprotocol/sdk ^1.29.0
• Node: >= 20
• License: MIT
• Author: Matthias Meyer (StudioMeyer)

Install

npm install mcp-stdio-shellguard

Or run the audit CLI directly without installing:

npx -y -p mcp-stdio-shellguard mcp-shellguard-audit scan ./src

What it gives you

Three layers, opt-in piecewise:

1. Library API — drop-in guardExec / guardSpawn you call from your own MCP server. Default-deny allowlist, sandbox profiles, replay window.
2. Audit CLI — mcp-shellguard-audit scan <path> walks the AST, reports 12 anti-patterns from LOW (no timeout) to CRITICAL (exec(\...${userInput}...`)`).
3. Reference MCP server — mcp-stdio-shellguard-demo exposes 8 tools so the MCP Inspector / Claude Desktop can drive the bundle directly.

Tools (reference server)

Sandbox profiles

Caller can tighten via timeoutMs / fdBudget per call. Caller cannot widen beyond the profile.

Trust tiers

Lift LOW → CRITICAL by registering the tool + setting argsPatterns + running through guardExec/guardSpawn (which always activate sandbox + replay).

Library quickstart

import {
  AllowlistRegistry,
  ReplayWindow,
  guardExec,
} from "mcp-stdio-shellguard";

const registry = new AllowlistRegistry();
const replay = new ReplayWindow();

registry.register({
  toolName: "git-log",
  executable: "/usr/bin/git",
  argsPatterns: ["^log$", "^--oneline$", "^-n$", "^\\d+$"],
  sandboxProfile: "strict",
});

const result = await guardExec(
  {
    toolName: "git-log",
    command: "/usr/bin/git",
    args: ["log", "--oneline", "-n", "10"],
  },
  { registry, replay },
);

console.log(result.stdout); // → commit lines
console.log(result.trustTier); // → "CRITICAL"
console.log(result.canonicalHash); // → 64-char SHA-256

Audit CLI

mcp-shellguard-audit scan ./src
mcp-shellguard-audit scan ./src --format sarif --output audit.sarif
mcp-shellguard-audit scan ./src --severity-floor HIGH    # CI gate

Exit codes:

• 0 clean (no findings at-or-above floor)
• 1 findings present
• 2 parse / IO errors

Anti-pattern library (12 rules)

The scanner resolves renamed child_process bindings before matching, so the dangerous shapes below are caught even when the call goes through an alias rather than a literal child_process.exec:

• const execAsync = promisify(exec); execAsync(...${x})
• import cp from "node:child_process"; cp.exec(...${x})
• const { exec: sh } = require("child_process"); sh(...${x})
• import { exec as run } from "node:child_process"; run(...)

Synchronous variants (spawnSync, execFileSync) share their async rules, and shell_true_option also fires on a string shell ({ shell: "/bin/sh" }) or a dynamic shell value — not just the literal { shell: true }. A promisify of a non-child_process function, a destructure off another module, and { shell: false } stay clean (no false positives).

Pragmas

• // shellguard:ignore-next-line — suppress one finding
• // shellguard:ignore-file — suppress whole file (rare; prefer per-line)

Why this exists

Ox-Security disclosed (2026-05) that 200k+ MCP stdio servers wrap child_process.exec with template literals carrying user input straight from LLM tool args. LiteLLM v1.83.6 was the canonical example (CVE patched in 1.83.7). This bundle is the defensive-security counterpart: a drop-in guard + scanner that closes the class. Inspired by AWS Linux seccomp + Chromium sandbox tiers.

See also

• HOOK_RECIPES.md — Claude Code hook recipes that auto-block dangerous tool calls
• CHANGELOG.md — release history
• Ox-Security MCP audit: https://venturebeat.com/security/200000-mcp-stdio-servers/
• LiteLLM CVE-2026-XXXX: https://github.com/BerriAI/litellm/security/advisories

License

MIT — Copyright (c) 2026 Matthias Meyer (StudioMeyer)