Exposed secrets
API keys, bearer tokens, private keys, passwords, database URLs, and high-entropy credentials stored in plain text.
Secret scanning guide → PRIVATE BY DESIGN · ZERO UPLOADS
Find the dangerous lines in your MCP config.
Paste a Claude, Cursor, Codex, Gemini, VS Code, Windsurf, Cline, or generic MCP configuration. Get an explainable risk score and exact remediation—without sending credentials to a server.
Runs locallyNo signupSecrets redactedOpen source
CONFIG AUDIT
MCP security scanner
Everything stays in this tab. Files are never uploaded.
Ready to inspect
We check structure, credentials, commands, packages, transports, and filesystem scope.
WHAT WE CHECK
Configuration risk, explained line by line.
A static config scan cannot prove a server is trustworthy. It can catch high-signal mistakes before the server starts with your agent's privileges.
Command execution
Shell interpreters, chained commands, command substitution, encoded PowerShell, curl-to-shell patterns, and privilege escalation.
Command injection guide →Supply-chain drift
Unpinned npx packages, automatic install flags, floating versions, Git URLs, and executable downloads.
Hardening checklist →Unsafe reach
Plain HTTP endpoints, credentials in URLs, broad filesystem roots, home-directory access, and wildcard paths.
Claude config security →BOUNDARIES MATTER
What this scan can—and cannot—prove
It can flag
- Risky values visible in JSON or TOML
- Commands likely to invoke a shell
- Unencrypted remote transports
- Overly broad paths and environment secrets
It cannot verify
- A package's source code or current owner
- Runtime tool descriptions and tool poisoning
- Server behavior after launch
- Whether every detected token is still active
Use this as a preflight check, then review the server source, publisher, requested capabilities, and runtime tool list.
FAQ
MCP configuration security questions
Is it safe to paste a live API key here?
The scanner runs entirely in your browser and makes no scan requests. Still, rotating a key that has been copied into an unknown environment is always the safest practice.
Does a clean score mean an MCP server is safe?
No. A configuration scan only evaluates visible setup risk. It cannot inspect changing tool descriptions, server source code, downstream APIs, or runtime responses.
Why is an unpinned npx package risky?
A floating package version can resolve to different code later. Pinning an exact reviewed version reduces unexpected supply-chain changes.
Which MCP clients are supported?
The scanner recognizes common JSON wrappers used by Claude Desktop, Cursor, VS Code, Windsurf, Cline, and Gemini, plus Codex TOML server blocks.
PURPOSE & METHODOLOGY
A focused tool, with the context to use it well
MCP Guard was created to find risky commands, exposed secrets, and unsafe patterns in MCP configuration. It is intended for developers and teams reviewing local MCP server configurations before use. The tool is free, requires no account, and is paired with original explanations so you can understand the result rather than copy an unexplained output.
How to use the result
Start with a sample or a non-sensitive copy of your data, run the check, and review every finding before changing a production project. Checks are paired with plain-English remediation guidance. The scanner is a defensive aid, not a certification, and users are encouraged to review commands and rotate any exposed credential. Specifications and software evolve, so confirm high-impact decisions against the linked primary documentation.
Privacy and limitations
MCP configuration text is processed in your browser wherever the tool page states that local processing is used. Avoid entering passwords, access tokens, personal data, or confidential material. Automated output may be incomplete because it cannot know your entire deployment, threat model, or organizational policy.
Independent and transparent
This is an independent developer resource, not an official certification service. Source code and issue tracking are available on GitHub. Read more about the project, review the privacy policy, or report an error.