Use Ctrl+P (or Cmd+P) to save as PDF. Back to paper
Before MCP, connecting an AI application to an external tool meant writing a custom adapter. A hundred tools meant a hundred adapters. Each new AI host meant rebuilding them all. The integration surface was M×N: every AI application multiplied by every external system.
The Model Context Protocol collapses that to M+N.
MCP is not a REST API. It is an open protocol — built on JSON-RPC 2.0 — that defines how LLM applications (hosts), their connectors (clients), and external services (servers) communicate in a standard way. Anthropic published the first version in November 2024. OpenAI adopted it by late 2025. The specification is maintained openly at modelcontextprotocol.io.
The protocol defines three roles:
The three primitives are:
Each session begins with capability negotiation. The client declares what it supports; the server declares what it offers. Nothing executes outside what both sides agree to.
Calling MCP "a way to add tools" undersells it. The architecture is a capability-boundary system. Each server connection is isolated — clients cannot cross into each other's servers. The protocol defines security boundaries at the transport layer, not as an afterthought in application code.
The transport is pluggable. Local programs communicate via stdio. Remote servers communicate via Streamable HTTP with OAuth-compatible authentication. The same JSON-RPC messages flow over both.
A developer writes one MCP server for a database. Every MCP-compatible host — any agent framework, any IDE, any AI app — can connect to it without modification. The integration is the protocol, not the adapter.
This is not a new idea. HTTP did this for the web. Language Server Protocol did it for editor integrations. MCP is doing it for AI tool connectivity — and unlike earlier attempts at AI tool standards, it arrived with host adoption from the start.
The M×N problem in AI tool integration is the same problem LSP solved in editor tooling in 2016. Understanding LSP's adoption arc tells you exactly what to expect from MCP.
1. Anthropic, 2024. "Introducing the Model Context Protocol." Anthropic News. https://www.anthropic.com/news/model-context-protocol
2. Model Context Protocol Authors, 2025. "Architecture — Model Context Protocol Specification (2025-06-18)." modelcontextprotocol.io. https://modelcontextprotocol.io/specification/2025-06-18/architecture
3. Model Context Protocol Authors, 2025. "Understanding MCP Servers." modelcontextprotocol.io. https://modelcontextprotocol.io/docs/learn/server-concepts
4. OpenAI, 2025. "Building MCP servers for ChatGPT Apps and API integrations." OpenAI Developer Documentation. https://developers.openai.com/api/docs/mcp
5. Hugging Face, 2025. "MCP Key Concepts and Architecture." Hugging Face Context Course, Unit 2. https://huggingface.co/learn/context-course/unit2/key-concepts