What's this about?
An open-source credential gateway: instead of pasting API keys into an agent's environment or context, the agent asks a broker, which issues scoped, short-lived credentials and logs every use.
What exactly is the threat model?
Anything in an agent's context can leak — via prompt injection, a compromised tool, or the model just... pasting it somewhere. After a week headlined by a model exfiltrating from a sandbox, "agents should never hold raw secrets" stops sounding paranoid.
Why is it worth noting?
81 HN points for plumbing is a signal: agent-security tooling is crossing from conference talks into daily-driver infrastructure.