MCP: the standard that made every tool speak the same language.
Before MCP, connecting an AI agent to a new tool meant writing custom integration code for every single one. Different authentication, different APIs, different error handling. MCP replaced all of that with one open protocol. Your agent speaks MCP once, and every MCP-compatible server understands it.
Four primitives, one protocol. Every MCP server exposes some combination of tools, resources, prompts, and a transport layer. The agent doesn't need to know the details.
A JSON-RPC protocol that lets an AI application discover and use anything.
MCP is an open, JSON-RPC 2.0 based standard. The AI application is the client. The MCP server is what exposes capabilities - tools, data, prompt templates. They communicate over a stateful session, which means the server remembers context across calls. This is different from REST, where every request starts from zero.
Anthropic originally published the specification, but MCP is now an open standard with servers available for everything from filesystems and databases to web search and code execution. Any AI application that speaks MCP can use any MCP server.
Everything an MCP server exposes falls into one of three categories.
Three primitives, each with a distinct job. No overlap, no ambiguity.
Tools
Callable functions the model invokes directly: query a database, send an email, search the web, execute code. Tools are the active primitives - the model decides when to call them and what arguments to pass.
Resources
Read-only data the model reads for context: file contents, database schemas, documentation pages. Resources are passive - the model reads them but doesn't change them.
Prompts
Pre-built instruction templates that tell the model how to work with specific tools and resources. Prompts are reusable patterns - write once, use across every MCP client.
Same protocol, two ways to ship the bytes.
Local
The agent spawns the MCP server as a subprocess and talks over standard input and output. Fast, synchronous, zero network overhead. Ideal for filesystem tools, local databases, and CLI wrappers.
Remote
The MCP server runs as a web service. The client connects over HTTP with server-sent events for streaming responses. Same primitives, different wire. Used for cloud APIs, remote databases, and shared services.
Before MCP, every tool was a custom engineering project.
Connecting an AI agent to a new data source meant writing authentication code, API wrappers, error handling, retry logic, and output parsers. Every tool was a separate project with its own bugs, its own maintenance burden, and its own way of failing.
MCP standardizes discovery, invocation, and error handling so the agent doesn't need to know anything tool-specific. Add a new MCP server, and every MCP client can use it immediately - zero per-tool engineering.
This is the same dynamic that made USB-C win over a drawer full of proprietary cables. One standard, universal compatibility, no per-device work. The protocol is the product.
"A protocol is just an agreement. MCP is the agreement that an AI agent and every tool it needs can finally share."
Every tool that speaks MCP works with every MCP client. The integration cost collapses to zero after the first one.
A stateful session means the server remembers. No stateless REST round-trips where the model has to reintroduce itself every call.