How to stop Claude Code and Cursor from leaking API keys
A system prompt that says "never leak secrets" does not stop the tool call. The check has to run before the command executes.
Where the leak happens
Claude Code and Cursor run shell and HTTP tools on your behalf. A command that prints an environment variable, posts a token, or copies a key into a request is already too late once the process starts. The useful control is a PreToolUse hook: inspect the command, then deny it.
Deny, do not warn-and-continue
ThumbGate treats a detected secret as a deny. The agent stops. It does not rewrite the same command with the secret still inside it. A warning on an internal policy, such as spend, is a warning. That warning is not permission to send a credential on the next call.
The same repository can also record a lesson. If that lesson already failed once, the repeat is a deny too.
What you run
npx thumbgate init
npx thumbgate gate-checkPass the tool name and the tool input to gate-check. A deny result ends the turn. Hosted Pro, which keeps shared lesson history and the dashboard, is separate from the public package.