Keep your API keys out of your AI agent: a credential pattern for MCP servers

A new article on Dev.to highlights the security risks of embedding sensitive API keys directly into AI agent configurations. When agents are granted direct access to MCP servers, they often hold long-lived credentials, creating a significant security vulnerability if the agent is compromised via prompt injection or malicious payloads. The author proposes a 'gateway' pattern, where a trusted intermediary process sits between the agent and the MCP server. This gateway manages credentials, enforces strict 'default-deny' access policies, and maintains a tamper-evident audit log of all tool calls. By moving the trust boundary, developers can ensure that agents only receive short-lived, scoped tokens rather than master keys. This approach minimizes the blast radius of potential compromises, ensuring that even if an agent is manipulated, the underlying system remains protected. The author demonstrates this pattern using the open-source tool Agenthof.