AI Tooling Configuration Scanner

Scan committed MCP server configs and AI coding-agent settings for inline API keys, unpinned auto-executed servers, and leaked internal endpoints.

How this scan works

Finds committed AI tooling configuration — .mcp.json, .cursor/mcp.json, .vscode/mcp.json, .claude/settings.json, .continue/config.json, .aider.conf.yml, Windsurf, Cline, Roo and Gemini equivalents
Flags API keys hardcoded in MCP server env blocks. These get committed constantly because the file reads as editor settings rather than as secrets, and the key is live all the same
Flags MCP servers launched with npx / uvx / bunx against an unversioned package: that fetches and executes whatever the name resolves to today, on every editor start, with your full local privileges and no lockfile recording what ran
Reads agent instruction files (CLAUDE.md, AGENTS.md, .cursorrules, copilot-instructions.md) for internal hostnames and private-range addresses — read verbatim by an agent with tool access, and by everyone else if the repo is public
Any credential found here is checked against its issuer like every other secret, so you learn whether it is still live