Agent Permission Audit

Your AI agent can probably do more than you think.

Paste your config and see exactly what it can reach.
Finds plaintext API keys, wide-open shell access and unpinned MCP servers in Claude Code, Cursor and VS Code configs. Runs 100% in your browser.

Audit my config

🔒 Nothing leaves your browser

Your config is checked right here, on your device. No uploads, no tracking, no account. This page even blocks itself from making network requests, so you can turn off your wifi after it loads and it still works.

Comments and trailing commas are fine.

Hardening Pack, $19

✓ Exact fix for every finding
✓ 3 locked-down config templates
✓ Key rotation runbook, 6 providers
✓ Pre-commit guard so it doesn't happen again

Also inside: a 12-question checklist for vetting an MCP server before you install it, and a calm step-by-step for when an agent does something bad.

Fix it, $19

One-time purchase. After you pay you land on your own Hardening Pack page. Bookmark it.

What we check

CriticalPermission prompts switched off (bypassPermissions)
CriticalPlaintext API keys: Anthropic, OpenAI, GitHub, Slack, AWS, Google, Stripe and JWT-style tokens
HighAny shell command allowed (bare Bash or Bash(*))
HighBlanket allow for curl, rm, sudo or ssh
HighFilesystem server that can see your whole disk or home folder
HighRemote MCP server over plain http://
Mediumnpx or uvx packages with no pinned version
MediumServers running from a temp folder
MediumNo deny rules for .env, *.pem or ~/.ssh
LowMore than 8 MCP servers switched on
LowDuplicate servers
LowUnscoped or unfamiliar package names

Score = 100 minus 30 per critical, 15 per high, 8 per medium and 3 per low, never below 0. A is 90+, B 80+, C 70+, D 60+, F under 60.

Questions

Does my config get sent anywhere?

No. The checker is JavaScript served with this page, and you can read all of it with your browser's View Source. The page tells your browser not to make any network requests at all (that's the connect-src 'none' rule in its security policy). We don't store what you paste, not even on your own device.

Is it safe to paste a file with real keys in it?

Yes, since nothing leaves the page. If you'd still rather not, keep the first 8 characters of each key and swap the rest for random letters. We only need the start to spot it, and we mask everything after the first 4 characters when we show it.

Which files can I check?

Claude Code settings (~/.claude/settings.json, .claude/settings.json, .claude/settings.local.json), MCP configs for Claude Desktop (claude_desktop_config.json), Claude Code (.mcp.json or ~/.claude.json) and Cursor (~/.cursor/mcp.json or .cursor/mcp.json), and VS Code (.vscode/mcp.json). We work out which one it is from what's inside.

Does a good score mean my agent is safe?

It means your config avoids the most common, most dangerous mistakes. It can't see what an MCP server actually does once it's running, or what's in your shell profile. The Hardening Pack's vetting checklist covers the parts a config check can't.

How do I get the fixes?

After you pay, Stripe sends you to your own Hardening Pack page: the same checker with every fix shown, plus the pack to download. There's no account. Bookmark the page to use it on other devices.

Why do unpinned packages matter so much?

npx -y some-package downloads whatever the newest version is, every time it starts. If that package is ever taken over, your next restart runs the new code with everything your agent can reach. Pinning a version means you only update when you choose to.