AI Pulse · Under the Radar

A small open-source project called OneCLI hit Hacker News this week and picked up real traction, 103 upvotes and 31 comments. The timing matters. It landed the same week OpenAI and Hugging Face dealt with a credential incident, and the core idea is simple: AI agents should never touch raw API keys or secrets at all.

OneCLI acts as a gateway. Instead of handing your agent a Stripe key or a database password, you give it a reference. When the agent needs to call Stripe, it asks the gateway, the gateway checks policy, and if allowed, the gateway makes the call on behalf of the agent. The agent never sees the actual secret. If the agent gets compromised or leaks its prompt, your keys stay safe.

This is not a novel idea in security circles, but the fact that it is gaining community attention now tells you something. The norm is shifting. A year ago, most people building agents stuffed credentials straight into environment variables or prompt context and called it done. Now, after a string of public leaks and close calls, the conversation is moving toward isolation by default. OneCLI is one answer. There will be others.

For anyone running agents in production, this is the direction you want to move. You do not need to adopt OneCLI specifically, but you do need a plan that keeps secrets out of agent memory and logs. That might mean a gateway, a broker, short-lived tokens, or policy enforcement at the infrastructure layer. The pattern is what matters, not the tool.

We build this kind of isolation into systems like the Omni Command Centre because once you have a dozen agents calling a dozen services, manual key management breaks down fast. The principle is the same whether you are running one agent or fifty: if the agent does not need to see the secret, do not show it.