Glossary
What is MCP (Model Context Protocol)?
The open standard, originated by Anthropic, that lets AI agents discover and use tools across the enterprise.
Definition
The Model Context Protocol (MCP) is an open standard, originally created by Anthropic, that defines how AI agents discover and use tools. Build an MCP server once, and every MCP-compatible agent — Claude, Cursor, Codex, Windsurf, ChatGPT, and others — can use it. Salesforce, DocuSign, Azure DevOps, GitHub, Notion, Slack, and most major enterprise platforms shipped MCP servers in 2026; Salesforce's Headless 360 announcement made every Salesforce capability available as an MCP tool. MCP is doing for AI agents what REST did for web services: standardizing the protocol layer that everything connects through.
Why it matters
The agentic enterprise needs a way for agents to talk to enterprise systems without writing custom integration code for every new agent-system pairing. Before MCP, every agent platform had its own tool-calling format. After MCP, the same MCP server works for any MCP-compatible client. This is the protocol-layer standardization that makes the agentic enterprise economically viable at scale.
But the protocol is only the easy part. Without governance — auth, audit trails, rate limits, lifecycle management — every MCP server an enterprise stands up is a new attack surface. See MuleSoft Agent Fabric (/glossary/agent-fabric/) for the productized governance layer and the governed API layer (/glossary/governed-api-layer/) for the broader pattern.
The core concepts
MCP server. Exposes tools, resources, and prompts that an MCP client (an agent) can discover and use.
MCP client. Typically an AI agent or coding assistant. Discovers MCP servers, calls tools, retrieves resources.
Tools. The discrete actions an MCP server exposes — each with a defined input/output schema agents can introspect.
Resources. Read-only data that an MCP client can retrieve.
Prompts. Reusable prompt templates an MCP server can offer to clients.
Transport layer. stdio, Server-Sent Events (HTTP), and WebSocket transports.
What it means for enterprise architects
MCP changes the integration math. Every API your business has built becomes potentially agent-accessible. But 'potentially' is the key word — production-grade agent access requires governance, observability, and lifecycle management. The cleanest architectural pattern is to expose MCP servers through a governed integration layer (MuleSoft) rather than directly from each system to each agent. That preserves the standardization win while preventing point-to-point integration sprawl from recurring at the agent layer. Wiring every agent directly to every system's MCP server is the protocol-shift version of the point-to-point integration mess MuleSoft was built to solve a decade ago. Different protocol, same architectural smell.
How Green Irony delivers MCP implementations
We build MCP servers as the operating infrastructure for our own AI Chief of Staff. Notion, Slack, Gmail, Calendar, DocuSign — all run through MCP for us, with rate limits, scoped credentials, and audit logs. We deliver the same pattern for clients via Run on Claude (/run-on-claude/) and SMB MuleSoft (/smb-mulesoft/) engagements. See our blog post MCP Is the New API (/blog/mcp-is-the-new-api/).