mcpguard — sample MCP configuration security review

Generated from a real scan of bad-config.json — a synthetic fixture, not a customer's system.
1 file scanned  ·  6 high 3 medium 9 total findings

Findings

SeverityRuleServerFieldWhat it meansHow to fix
HIGH MCP001 fs args Package '@modelcontextprotocol/server-filesystem' is not pinned to an exact version (found 'latest') and is auto-installed with -y/--yes. Pin an exact version, e.g. @modelcontextprotocol/server-filesystem@2026.7.1.
HIGH MCP002 fs env.GITHUB_TOKEN GitHub token assigned to 'GITHUB_TOKEN'. Reference an environment variable instead, e.g. ${GITHUB_TOKEN} or env:GITHUB_TOKEN.
HIGH MCP002 shell env.API_KEY AWS access key assigned to 'API_KEY'. Reference an environment variable instead, e.g. ${GITHUB_TOKEN} or env:GITHUB_TOKEN.
HIGH MCP004 shell command Command delegates to a shell interpreter (-c / /c / -enc) or uses eval/exec. Invoke the target binary directly with a structured argument array; avoid shell -c wrappers.
HIGH MCP005 remote args[0] Security-disabling flag '--no-sandbox' is present. Remove the flag. If a sandbox conflict occurs, fix the underlying cause rather than disabling the control.
HIGH MCP008 remote args[2] OpenAI-style key embedded in command-line argument. Pass the credential via an environment variable reference instead of a literal argument.
MEDIUM MCP003 fs args[2] Filesystem-style server is granted broad root '/'. Scope the server to the smallest project directory it actually needs.
MEDIUM MCP006 remote url Remote MCP server uses cleartext HTTP to a non-loopback host. Use an https:// URL (or a loopback address for strictly local development).
MEDIUM MCP007 fs allowedTools[0] Wildcard tool permission at 'allowedTools[0]'. List only the specific tool names the workflow requires.
Scope of this review. mcpguard performs static inspection of MCP configuration JSON only. It does not execute commands, connect to servers, test running systems, or review server source code. A clean result means none of the eight documented patterns were found — not that the configuration is secure. Secret values shown in findings are masked.

What a paid review includes

  1. Run mcpguard across the supplied MCP config files.
  2. Written finding-by-finding report with severity, field, rationale, and concrete fix.
  3. Corrected example config (diff) with secrets replaced by environment-variable references.
  4. Optional SARIF file for GitHub code-scanning integration.
  5. One recheck after fixes are applied.

Source and test suite: github.com/sidsri14/mcpguard