HeimWall
Catch secrets before they leak into Cursor & Claude
47 followers
Catch secrets before they leak into Cursor & Claude
47 followers
HeimWall catches leaked secrets, credentials, and PII the moment they're about to reach AI coding assistants like Cursor, Claude Code, and Copilot. A lightweight macOS app, fully on-device: 47 hand-written rules flag leaks in real time. Your prompts never leave your Mac. No content stored, no account, no signup. Free for individual engineers. Next up: a team dashboard showing security leads leak trends without exposing what anyone typed. Signal, not content. Design partners welcome.
would love to see a quick test mode where i can paste in a sample prompt and see exactly which rules flagged it, basically a way to tune the 47 rules for my own workflows without needing to trigger real leaks in my actual coding sessions
@ramazan793230 You can do a rough version of this today: the clipboard guard scans whatever you copy, so copying a sample key lights up the feed with the exact rule that matched, no coding session involved.
A proper playground panel inside the app is a lovely idea though. The engine runs right there on-device, so a type-and-see-what-fires box is very doable, and honestly it would double as the best onboarding demo. Adding it to the list.
The on-device approach is a big deal for trust. One thing that would help me roll this out to my team is a dry-run mode that lets me see which of my recent prompts would have been flagged, so I can fix habits without the awkward retroactive alerts.
@hafizeewzm Small confession: we can't do retroactive, for exactly the reason you like us. We never store prompts, so there's nothing to go back and scan. That tradeoff is deliberate.
The good news is the app is already dry-run by design. It never blocks, it only flags locally, so a normal week of use gives you that exact report: your own private feed of what would have leaked, visible to nobody but you. And with the playground panel another commenter suggested today, you'll be able to paste old prompts in and see what fires.
honestly this looks really useful for anyone paranoid about pasting real keys into copilot by accident. one thing i'd love is a quick "why was this flagged" popover when a rule triggers, basically a one-liner explaining which pattern matched so i can learn what to scrub next time. would make the whole thing way less mysterious
@serapkaratop Love this. Half of it is already there, every catch in the feed shows which rule matched and its category. The plain-language one-liner is a great addition though, and it's the nice part of running a deterministic rule engine: every flag has an exact answer to why, nothing mysterious under the hood. Adding it to the list.
love that this stays fully on-device, that was basically the deciding factor for me. one thing though, would be great if you could add a quick toggle to whitelist specific projects or repos, like a dotfile in the project root, so it doesn't flag test fixtures or seed data as real leaks. right now i imagine anyone with mock credentials in their codebase will get noise.
@mahmutjldv Thanks! One mental-model note, we scan the prompt box itself, not your files, so the app doesn't know which repo a paste came from. That makes a project dotfile tricky, but an allowlist in some form is a fair ask and it's noted.
On noise, the validators already reject placeholder-shaped values: AWS EXAMPLE keys, your-token style strings, values that look like code rather than secrets. Fixtures with authentic-format values will still flag, though arguably a mock key indistinguishable from a real one deserves the ping.
Ran it for an afternoon while pushing code through Cursor and it actually flagged a stray API key I had no business pasting in. The "fully on-device" bit sold me, no setup, no signup, just a quiet macOS app doing its job in the background.
@saliha464180 Thanks Saliha, that catch is exactly the point and that is why we are building this. Many times I realized afterwards that I shared credentials or keys with agents, so these kind of applications are needed to prevent this before it happens. Also, doing it fully on-device is the part I'm proudest of. If it ever misses one or flags something it shouldn't, please let us know.
Dial
this is one of those problems that's obvious in hindsight but nobody built for until now, pasting a .env file into a chat window to get context is such a common shortcut. how does detection work for secrets that don't match a known pattern, like an internal company token or a randomly generated key that isn't an AWS/Stripe style prefix? that feels like the harder 80% of this problem.