
The broker
An AI coding agent that can run commands can print any credential in its reach. The usual answer is a blocklist of ways to print an environment, and that list is not enumerable: an interpreter one-liner, a script's error output, a tool logging its own config, a test fixture. The fix is absence.
We offer:
- Identity read from the kernel per connection, never from a header
- Method and path checked against a per-service allowlist
- A rate limit per caller, per key
- The credential injected on the way out and stripped from anything coming back
- One audit line per call, refusals included



