Credential Broker

Simple Definition

A credential broker is a trusted go-between that handles logins and secrets for an AI agent. The agent asks to authenticate, the broker checks whether that is allowed, and the broker performs or authorizes the login, so the raw password or key never passes through the model.

Plain-English Explanation

The principle is simple: an agent may need the ability to authenticate, but it does not need to know the secret. Anything that enters an agent’s context can leak. It can be written to logs, saved in memory, repeated in an output, or pulled out by a prompt injection hidden in a web page. A broker removes the secret from that path.

In AI-agent systems the term is used loosely, and implementations differ. Some brokers fill a login form directly in the agent’s browser. Others swap a placeholder for the real API key only when a request goes to an approved destination. Others issue short-lived tokens. What they share is that the model works with a reference or a session, never the secret itself.

Example

An agent needs to log in to a utility account to download a bill.

1. The agent reaches the login page and asks the broker for the right credential.
2. The broker checks its policy: is this agent allowed to use this login for this task?
3. If the policy requires it, you get a prompt to approve.
4. The broker fills in the login, and the agent continues with an authenticated session.
5. The password never appears in the agent’s context.

1Password for Agents works this way for personal agents. NVIDIA OpenShell does something similar for developer agents, injecting credentials only into requests bound for approved endpoints.

How It Differs From Related Terms

A password manager or secret manager stores credentials. A credential broker mediates their use, often on top of one of those stores. The overlap is real: some password managers now include broker features. An authorization gate decides whether an action may proceed; a broker decides whether and how a credential may be used, which is often one of the things a gate checks. Brokering protects the secret but does not limit what the agent does after it is logged in. That still needs an approval policy and narrow permissions.

Related Terms

• Non-Human Identity, the identities brokers often serve
• Authorization Gate, the permission check before an action
• Approval Policy, the rules for when a login needs your approval
• AI Permission Hygiene, limiting what each login can reach
• Agent Sandbox, the isolated environment brokers often work with