MCP (Model Context Protocol)
The standard protocol for connecting LLMs to external tools. Not a package you install: an interoperability specification. Concrete MCP servers are the actual tools. Before adding an MCP: weigh the token cost. Each server loads its tool descriptions into the agent context, and the agent has to pick among them on every request: the official GitHub server groups its tools into about twenty families, five of them on by default, and every family you add weighs on the window. Three cases to tell apart. Two to five tools: declare them to the agent directly, a plain tool call is enough. A deterministic action, such as reading a customer record or checking an order: do not even involve the model, call the API from your code. MCP earns its place when several agents need the same integration, such as the internal database: you write one server, and every agent in the company plugs into it without rewriting the connection. The arithmetic behind it, laid out by [Bunty Shah](https://www.instagram.com/reel/Dd_En54A266/): two applications each talking to ten services means twenty integrations to write and maintain (authentication, calls, errors, retries); with MCP it is two clients plus ten servers, and next year's application only adds one client. Two misreadings to avoid. The model never calls the API itself: it says which action it wants, and your application runs it. And MCP does not replace the API: the MCP server sits in front of it as a translator, calls it the usual way, and tells the client which actions are available (discovery). The model picks from that list, bounded by what your application allows.
Strengths
- Open Anthropic standard, adopted by Claude, Cursor, Windsurf, Gemini CLI
- Clear client/server architecture: one MCP server = one tool, one namespace
- Multiple transports: stdio (local), HTTP SSE (remote), WebSockets (real-time)
- Explosive community in 2026: +10k public MCP servers listed
Limitations
- Large security attack surface: prompt injection via tool descriptions
- No authentication standard yet, each server handles its own auth
- Hard to debug: MCP errors barely readable without dedicated tooling
- Token cost: every active MCP server consumes context to expose its tools
Best for
- AI agents requiring access to APIs, files, databases
- Integrating existing tools (GitHub, Jira, Slack) into an LLM workflow