# MCP (Model Context Protocol)

> MCP is an open protocol that lets AI agents connect to external tools and data through servers — the standard way to extend coding agents like Claude Code.

Source: https://tokenade.net/en/glossary/mcp · Updated: 2026-10-07

## What is MCP (Model Context Protocol)?

MCP, the Model Context Protocol, is an open standard that lets AI agents talk to external tools and data sources through small programs called MCP servers. Instead of each agent inventing its own integration format, an agent that "speaks MCP" can connect to any MCP server — for filesystem access, GitHub, a browser, documentation, and more — using one common protocol.

In practice, you add MCP servers to a coding agent like Claude Code, Cursor or Cline, and the agent gains those servers' tools. Each server advertises its tools through a manifest the agent loads.

Anthropic created MCP and, in December 2025, donated it to the Agentic AI Foundation, a directed fund under the Linux Foundation co-founded with Block and OpenAI. As of October 2026 the current specification revision is 2026-07-28. Its main changes versus 2025-11-25: protocol-level sessions and the `Mcp-Session-Id` header are gone from the Streamable HTTP transport (so `tools/list` no longer varies per connection), long-running tasks moved into an official extension, and features now follow a lifecycle with a minimum twelve-month deprecation window.

## Why MCP matters in 2026

MCP matters because it's become the de-facto way to extend coding agents — but it has a token cost worth understanding. The tool definitions from every connected server are sent to the model as part of the context, and re-sent on each turn, so adding many servers quietly inflates token usage whether or not you use those tools. The fix is loading tools lazily and pruning servers you don't use; see [Best MCP servers for Claude Code](https://tokenade.net/en/articles/best-mcp-servers-for-claude-code).

Clients have started to do the lazy part for you. As of October 2026, Claude Code's docs say MCP tool search is on by default: tool definitions are deferred and discovered on demand instead of all being loaded upfront. The catch is in the fine print — Claude Code falls back to loading every tool upfront when `ANTHROPIC_BASE_URL` points to a non-first-party host, on some cloud deployments, and when `ENABLE_TOOL_SEARCH=false`. Other clients vary, so check what yours actually sends before assuming the manifest tax is gone.

## When MCP isn't the right tool

- **For tools you rarely call** — a server you almost never use still taxes every turn; remove it.
- **For simple local actions** — sometimes a direct command is cheaper than routing through a tool manifest.

## See also

- [Glossary: Token](https://tokenade.net/en/glossary/token)
- [Glossary: Context window](https://tokenade.net/en/glossary/context-window)
- [Article: Best MCP servers for Claude Code](https://tokenade.net/en/articles/best-mcp-servers-for-claude-code)
- [Article: Claude Code MCP — how to set it up](https://tokenade.net/en/articles/claude-code-mcp-setup)
- [Article: MCP servers — what they are and what they cost](https://tokenade.net/en/articles/mcp-servers-explained)