Simon Willison’s new llm-keys-ui is a small answer to a common remote-agent problem. He uses coding agents from his phone to operate machines elsewhere, but does not want to paste API keys into their chat history. The command launches a local interface—reachable through a private network such as Tailscale—that accepts and stores a key without displaying it to the agent. The agent can later request the credential by name. Willison’s walkthrough and code
The artifact is compact enough to inspect, which is its appeal. It narrows accidental disclosure in the transcript and keeps a human in the credential-entry step. Its protection is deliberately narrow; host compromise, process access, logs and key rotation still matter. As a piece of agent plumbing, it gives “ask the human for a secret” a cleaner implementation than “please paste it here.”