@rn-security/mcp

AI-assisted security review for React Native projects — without an API key, a vendor, or a source upload.

Part of the React Native Security Toolkit.

claude mcp add rn-security -- npx -y @rn-security/mcp

Then ask your model to audit the project. It receives the toolkit's findings: severity, confidence, evidence, remediation, CWE / MASVS / MASWE references with official titles, and the MASTG tests that verify a fix.

Why this way

• No vendor. It speaks the Model Context Protocol, so the model is whichever one you already use — Claude Code, Claude Desktop, an editor, a local model behind an MCP client.
• No source upload. The server talks over stdio to a client on your machine and sends nothing anywhere.
• No API key. There is no credential for this tool to store or leak.
• Non-authoritative by construction. Every finding the model sees came from a deterministic rule with a stable identifier. A model can interpret a report; it cannot write one or change a severity.

Tools

| Tool | Returns | | ---------------------------- | ---------------------------------------------------------------------------------------- | | security_audit | Findings with severity, evidence, remediation, standards references and a coverage block | | security_rules | Every rule, with categories, standards and documentation paths | | security_rule_details | One rule in full, including the MASTG tests that verify a fix | | security_runtime_readiness | Whether the project declares what the on-device checks depend on |

All read-only. Nothing writes a file, installs a package, executes project code, or opens a socket.

What it refuses to do

Read outside your project. The model chooses every tool argument, and can be talked into choosing anything — including by the repository it is auditing. Paths are confined to the server's root.

Let your code give it orders. Everything quoted from your repository is labelled as untrusted data, and injection attempts are reported rather than obeyed — and never rewritten, because stripping the text would make the scanner lie about what a file contains.

Any MCP client

{
  "mcpServers": {
    "rn-security": { "command": "npx", "args": ["-y", "@rn-security/mcp"] }
  }
}

MCP 2025-06-18 over stdio, implemented directly — no protocol SDK, no transitive dependencies.

Documentation

Full MCP documentation

Licence

MIT © Muhammad Ahmad