MCP (Model Context Protocol) is an open protocol that standardizes how AI agents connect to external tools, replacing bespoke integrations with a single interface. The client-server handshake exposes tools via name, description, and JSON Schema, with the host executing requests rather than the model directly. Because agents that can act can be tricked via prompt injection (OWASP LLM01), the security approach is bounding what the model is allowed to do at the host level: allowlists, spend caps, per-action limits, human approval for irreversible actions, and scoped credentials. The piece also covers how MCP is becoming infrastructure for onchain agents, with Coinbase's x402 payment protocol and agentic wallets enabling agents to move real value within policy-enforced limits.
