Model Context Protocol (MCP) Explained: How AI Agents Connect to Any Tool

What happened

I'll pull the source article and the post-writing rules so the paragraphs stay factual and match the required structure.The source page was thin. I’ll pull the official MCP docs so the paragraphs stay specific and don’t invent versions or dates.Stika Studio published a walkthrough titled What Is Model Context Protocol (MCP)? Complete Guide 2026 at stikastudio.com/blog/what-is-model-context-protocol/. The piece, filed under AI Agents, walks through what MCP is, how MCP servers expose tools and resources, and why the protocol is meant to replace one-off LLM integrations, and it ships with code examples. The same URL landed on Hacker News as item 49297428 in the HN AI Agents feed, sitting at 1 point with 0 comments. That listing is a pointer, not a launch. Anthropic already open-sourced MCP on November 25, 2024 as a universal protocol for connecting AI assistants to the systems where data lives, including content repositories, business tools, and development environments. The stated goal was to stop frontier models from being trapped behind information silos and custom connectors.

MCP is a host, client, and server split, not a model plugin. An MCP host such as Claude Desktop, Claude Code, Visual Studio Code, Cursor, or ChatGPT creates one MCP client for each connected server. Each client keeps a dedicated JSON-RPC 2.0 channel to that server. The data layer defines discovery, capability exchange, and three server primitives: tools, which are executable functions the model can invoke (file operations, API calls, database queries); resources, which are readable context such as file contents or records; and prompts, which are reusable templates. Clients discover tools with tools/list and invoke them with tools/call; resources use list and read paths. The transport layer is separate. Local servers typically speak stdio to a single process on the same machine. Remote servers typically speak Streamable HTTP, using HTTP POST for client-to-server messages and optional Server-Sent Events for streaming, with OAuth recommended for tokens. The language model never talks to Slack or Postgres directly. The host injects discovered tool schemas into the model, routes a tools/call to the matching server, and folds the result back into the conversation.

The technical detail

That split is what changes the engineering work. Before MCP, every new data source needed its own function schema, auth path, and host-specific adapter. Anthropic described that as an N by M integration problem: N models or agents times M tools. A team that exposes a GitHub repo, a Postgres schema, and an internal FAQ as three MCP servers can attach those same servers to any host that speaks the protocol. Official docs compare this to a USB-C port for AI applications. Builders stop rewriting the same Slack or Git connector for Claude, ChatGPT, and Cursor. They write one server that advertises a natural-language description and a JSON Schema for each tool, then let the host handle discovery and invocation. The protocol does not dictate how the host uses the returned context. That is the host’s job.

Why it matters for builders

The market already treated that as a shared socket rather than an Anthropic-only feature. MCP was created at Anthropic by David Soria Parra and Justin Spahr-Summers. At launch Anthropic shipped the specification and SDKs, local MCP server support in the Claude Desktop apps, and an open-source repository of servers, including pre-built servers for Google Drive, Slack, GitHub, Git, Postgres, and Puppeteer. Early adopters named in the launch post included Block and Apollo. Development-tool companies including Zed, Replit, Codeium, and Sourcegraph were already wiring MCP so coding agents could pull project context instead of guessing from an open buffer. OpenAI later adopted the protocol across products including ChatGPT. Google DeepMind also moved to support it. In December 2025 Anthropic donated MCP to the Agentic AI Foundation, a directed fund under the Linux Foundation co-founded by Anthropic, Block, and OpenAI. Client coverage now includes Claude, ChatGPT, Visual Studio Code, and Cursor, which is the practical test of whether “build once” is real.

Market and competitive context

The practical next step is to pick a surface and treat MCP as the public interface, not as a vendor demo. If you own a product API that agents should call, ship an MCP server that lists a small set of well-described tools instead of another bespoke function-calling wrapper. If you are building an agent or IDE, consume servers through tools/list and tools/call rather than hard-coding vendor SDKs. Watch two implementation details. First, transport: stdio is the local default; Streamable HTTP plus OAuth is the remote production path Anthropic said it would back with developer toolkits for Claude for Work. Second, host coverage: a server that only works in Claude Desktop is a connector, not a protocol win. Official docs and SDKs live at modelcontextprotocol.io and github.com/modelcontextprotocol. The Stika Studio guide is useful as a code-level walkthrough of servers exposing tools and resources; the specification is the thing to pin against.

What to watch next

Risks sit in the same place the value does. Security researchers have shown prompt injection against MCP setups and poisoned tools that can exfiltrate data through other connected tools on the same host. A host that attaches many servers expands that blast radius, because the model sees a union of tool descriptions and can be steered into calling the wrong one. MCP does not define authorization policy beyond what the transport and the server implement, so a filesystem or database server is as privileged as the process that runs it. Open questions include how hosts should sandbox untrusted community servers, how tool lists stay small enough for the model to choose well when dozens of servers are attached, and how remote production auth should look once OAuth is the default. Related prior art is specific. OpenAI function calling and the ChatGPT plugin framework solved similar problems with vendor-specific connectors. OpenAPI describes REST APIs for humans and HTTP clients; MCP describes model-facing primitives with natural-language descriptions. The Language Server Protocol is the closest architectural ancestor: one language server, many editors. Agent2Agent is a different protocol, aimed at agent-to-agent communication rather than model-to-tool context. MCP Apps later extended the base spec so servers can deliver interactive UIs, not only text and structured data, which is a sign the protocol is still being stretched past the original tool-and-resource loop.