@khushwant.r/promptguard

PromptGuard

A secret firewall for your AI prompts.

When you're deep in a coding session, it's easy to paste a real API key, token, or .env value straight into a Cursor or Claude Code prompt — "here's my key, go run this for me." The moment you hit send, that secret leaves your machine and lands in a chat history you don't control.

PromptGuard sits in front of that send button. It scans every prompt locally, and if it spots something that looks like a secret, it blocks the prompt before it ever hits the network. The secret never leaves your laptop.

How it works

PromptGuard isn't an extension — extensions can't read the chat box. Instead it plugs into the prompt hooks both tools already expose:

• Cursor — beforeSubmitPrompt in ~/.cursor/hooks.json
• Claude Code — UserPromptSubmit in ~/.claude/settings.json

Each time you submit a prompt, the tool pipes it to PromptGuard. If it's clean, the prompt goes through. If it contains a secret, PromptGuard blocks it and shows you a message telling you what it caught.

Install

You'll need Node.js 18+ on your PATH.

npx @khushwant.r/promptguard install

That wires up the hook for both Cursor and Claude Code. Restart Cursor (or reload Claude Code) so it picks up the new hook.

Want just one tool, or a stricter setup?

npx @khushwant.r/promptguard install --cursor          # Cursor only
npx @khushwant.r/promptguard install --claude          # Claude Code only
npx @khushwant.r/promptguard install --project         # this repo only (.cursor / .claude)
npx @khushwant.r/promptguard install --strict          # fail closed: if the hook breaks, block
npx @khushwant.r/promptguard install --aggressive      # also flag values assigned to secret-ish names

Remove it just as easily:

npx @khushwant.r/promptguard uninstall

Try it

Send a prompt with a fake key in it, e.g.:

run this for me with AKIAIOSFODNN7EXAMPLE

PromptGuard will stop the submission and tell you a secret was detected. Delete the key and send again — it goes straight through.

What it catches

• Known provider credentials — AWS keys, OpenAI / Anthropic keys, GitHub & GitLab tokens, Stripe, Google API keys & OAuth secrets, Slack tokens & webhooks, SendGrid, Twilio, npm tokens, JWTs, private key blocks, Azure DevOps PATs, DigitalOcean & Open VSX tokens, and more.
• Anything that looks like a key — long, random, high-entropy strings, even if they don't match a known provider format.
• Secret-ish assignments (with --aggressive) — values handed to names like apiKey, password, token, clientSecret, etc.

It deliberately ignores obvious non-secrets — placeholders (your_api_key), import paths (@angular/core), UUIDs, URLs, emails, and file paths — so it stays out of your way on normal prompts.

Safety & privacy

• Everything runs locally. PromptGuard makes no network calls and stores nothing. Your prompts and secrets stay on your machine.
• By default it fails open — if the hook ever errors, your prompt still goes through so you're never stuck. Use --strict to fail closed instead (block on error), which is the safer choice for security-critical setups.
• When it does block, it only ever shows a masked preview of the secret (AKIA…LE), never the whole thing.

Commands

Command	What it does
promptguard scan	Reads a hook payload on stdin and allows/blocks it. Used by the hook.
promptguard install	Registers the prompt hook in Cursor and/or Claude Code.
promptguard uninstall	Removes only PromptGuard's hook entries.
promptguard --version	Prints the version.

Support

If PromptGuard stopped a secret from leaking into a chat log, consider giving it a on GitHub — it genuinely helps.

Built by Khushwant R.