Promptwall

When you work with AI coding tools, you paste things in all day: an error log, a config file, a chunk of code. Sooner or later an API key or a customer's email goes in with it. Nobody does it on purpose, and once the prompt is sent, you can't take it back. Most companies know this happens and have no way to stop it.

I wanted to catch the secret on the developer's own machine, in the moment before the prompt is sent. The company sets the rules, the developer barely notices it is there, and Promptwall never keeps a copy of what anyone typed.

Promptwall runs in the background and hooks into Claude Code, Codex, Gemini CLI and Copilot CLI. Every prompt is checked before it goes out, and most go straight through. If something looks risky, Promptwall does one of three things: it warns you and waits for your OK, it removes the secret and hands you a clean version to send, or it blocks the prompt. Security teams set the rules in a web dashboard, and developers can see what was caught, and why, in a small desktop app.

1. The background agent that checks each prompt before it is sent
2. Connections to Claude Code, Codex, Gemini CLI and Copilot CLI, using the hooks each tool officially supports
3. A rules engine and dashboard where a security team decides what is allowed, what gets a warning, what gets cleaned and what gets blocked
4. A safe default: if anything fails, the prompt is held back. Nothing slips through because something broke.
5. A desktop app where developers see what was caught, approve warnings and turn integrations on or off

• Secrets are caught on the laptop, before any AI tool sees them
• Prompt text and secret values are never saved. The logs record what happened, not what was typed.
• It keeps working offline with the last set of rules it was given
• Claude Code support is tested against the real app. The other tools are marked as early support until they are too.