MCP Security KitMCP config security checker
Paste the MCP config from Claude Desktop, Cursor, VS Code or Windsurf. You'll see hard-coded secrets, unpinned packages, shell and file-write access, remote servers without encryption or credentials, risky container settings and hidden instructions, with a fix for each.
Your config stays in your browser. The check runs on this page without uploading your config. The page may request static files from this site and contact its cookieless page-view counter. If optional Google analytics is configured, the page also loads the Google tag and contacts the configured analytics services.
Where to find it: Claude Desktop claude_desktop_config.json · Cursor ~/.cursor/mcp.json · VS Code .vscode/mcp.json · Windsurf ~/.codeium/windsurf/mcp_config.json
What this checks
- Secrets written into the file: API keys, tokens and passwords in
env, headers, arguments or URLs. Values are shown partly hidden. - Unpinned packages and images:
npx,uvxanddockerservers that fetch whatever version is newest on every start. - File access: filesystem servers that can write, edit and move files, and ones given
/, your home folder or a whole drive. - Shell access: servers launched through a shell or built to run shell commands, and tools set to run without asking.
- Remote servers:
http://endpoints off your machine, and remote servers with no credentials configured. - Container settings:
--privileged, host networking, the Docker socket and broad host mounts. - Tool poisoning in the config: hidden instructions in descriptions or notes, references to secret files and invisible characters.
Every check explained, with before-and-after configs.
Reviewing a server before you add it?
A server's own tool descriptions aren't in your config. Paste its tool list into MCP Server Check to find tool poisoning and tool shadowing before install.
More free checks for AI developers
- AI SDK dependency scanner: paste a package.json, requirements.txt, pyproject.toml or go.mod and see which AI SDKs, agent frameworks and MCP libraries a project pulls in.
- AI API key scanner: paste a .env, log or ticket and find OpenAI, Anthropic, Google, Hugging Face, AWS and other keys before they spread.
Questions
Is my config sent anywhere?
No. Files are read on your computer and nothing you paste or upload is sent anywhere. The page may request static files from this site and contact Cloudflare's cookieless page-view counter (connect-src 'self' https://cloudflareinsights.com), which never receives your config. You can disconnect from the internet after the page loads and it still works.
Which clients does it understand?
Configs with an mcpServers block (Claude Desktop, Cursor, Windsurf and most others), VS Code's servers block including its comments and trailing commas, and Zed's context_servers block.
Can it see a server's tool descriptions?
Only text that's in your config file. A server's own tool descriptions come from the server when it runs; to review those before installing, paste the server's tool list into MCP Server Check.
Is a clean result a guarantee?
No. These are pattern checks on the file, a quick first pass rather than a full review.
Sources
- Model Context Protocol: Security best practices
- Model Context Protocol specification (2025-06-18): Tools
- Checked 28 September 2026.