Free tool · in the browser

Is your Claude Code / MCP setup dangerous?

Paste a Claude Code settings.json, an MCP config (.mcp.json, claude_desktop_config.json, or Cursor mcp.json), or a CLI command. You get a score from 0 to 100, a level, and a plain-language fix for each finding.

Config or command

Your config is checked in your browser. It isn't uploaded, stored or sent to Vulnify or anyone else, and it isn't put in the page address. If you allow analytics, Cloudflare Web Analytics (no cookies) counts the page visit, never what you paste. Remove secrets before pasting anyway.

Add runtime approval to your agent

A config file shows what the agent is allowed to do. Vulnify answers allow, review, or block when a tool call is about to run.

Connect Vulnify in Claude Start free

Or audit from the terminal:

CLI

npx @vulnify/cli audit

What the score looks at

Skip-permissions flags, bypassPermissions, and CLAUDE_CODE_EVEN_MORE_DANGEROUSLY_SKIP_PERMISSIONS in the shell that launches Claude.
Wildcard shell allows such as Bash(*), plus rm, curl piped into a shell, and sudo.
A missing deny list, and read or write access to .env and ~/.ssh.
MCP servers launched with npx or uvx and no pinned version.
Filesystem servers rooted at /, a home directory, or a secret path.
Tokens written into env or headers, and web fetch with no domain limit.

Deny rules are read before allow rules. An allow rule cannot open a hole in a deny rule. This page only sees the text you paste.