Flashcard: why connect agents through MCP?
MCP is an open protocol from Anthropic (late 2024) that standardizes how LLM agents connect to tools, data, and prompts. It collapses N-by-M custom integrations into N-plus-M shared interfaces.
Imagine every brand of phone needing its own special charger, so a Samsung charger does not work on an iPhone and an iPhone charger does not work on a Pixel. That is the world before MCP: every agent framework had its own way to talk to tools, so every tool had to build a separate adapter for every framework. MCP is like USB-C for AI agents: one standard plug. A tool builder makes one MCP server, and any MCP-compatible client (Claude, Cursor, LangGraph) can use it. An agent builder makes one MCP client, and the agent can use any MCP-compatible tool. The number of integrations to build drops from N-times-M to N-plus-M, which is why MCP spread fast across 2025-2026.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
MCP (Model Context Protocol) is an open standard introduced by Anthropic in late 2024 for how LLM agents connect to tools, data sources, and prompt providers. By 2026 it is the de facto interop layer across the major agent runtimes, and any new agent framework that speaks MCP gains immediate access to a large catalog of public servers.
The motivation is a familiar one in software systems: collapse a quadratic integration problem into a linear one. Before MCP, every agent framework had its own tool format, and every tool provider had to build a separate adapter per framework. With N frameworks and M tools, integration work scaled as N-times-M, which is wasteful. MCP gives both sides a common target so the work scales as N-plus-M.
This explanation defines MCP precisely, names its three capability primitives, contrasts it with provider-native function calling, and explains why the protocol took off so quickly.
What MCP is
MCP is an open protocol for connecting LLM agents to external capabilities. The specification is published openly (modelcontextprotocol.io), reference implementations are open source, and the protocol is transport-agnostic.
The core abstractions are simple. An MCP server is a process that exposes capabilities. An MCP client is a process that consumes them. The two communicate over JSON-RPC, either via stdio (for local servers running on the same machine as the client) or over HTTP with Server-Sent Events (for remote servers). At connection time, the client and server perform a handshake where the server advertises its capabilities and version.
The three capability primitives are tools, resources, and prompts. Each addresses a different integration surface.
Tools are model-callable functions with JSON Schema-typed inputs and outputs. A tool might be 'search_github_issues' with inputs (repo, query) and output (a list of issues). The model emits a structured call to the tool, the server executes it, and the result is returned to the model.
Resources are addressable data items the model or user can read. A resource is identified by a URI and resolves to text or binary content. A filesystem MCP server might expose every file under a directory as a resource; a database MCP server might expose every table as a resource. Resources differ from tools in that they are addressable (you can list and reference them) rather than invocable.
Prompts are reusable prompt templates that the user or agent can invoke from a UI menu. A 'summarise pull request' prompt might prefill a structured query for the model. Prompts are how user-facing actions surface in MCP-aware UIs like Claude desktop and Cursor.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
Real products, models, and research that use this idea.
- Anthropic's Claude desktop app ships with MCP support out of the box; users add a filesystem or GitHub MCP server and Claude immediately gains those capabilities.
- Cursor IDE uses MCP to expose its codebase tools to the agent, and lets users add third-party MCP servers for custom integrations.
What an interviewer would ask next. Try answering before peeking at the approach.
QHow is MCP different from provider-native function calling?
Function calling (OpenAI tools, Anthropic tool_use, Google function_declarations) is how the model emits structured tool-call JSON during inference. MCP is how the runtime advertises tool capabilities, transports the call to a server process, and returns the result. They are complementary: function calling is model-side, MCP is system-side. A typical 2026 agent uses both.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Calling MCP an Anthropic-only protocol. It was introduced by Anthropic but is open and now used across Claude, Cursor, OpenAI Agents SDK, and most agent frameworks. The open-protocol property is what made it spread.
60 second bullets to scan on the way to the call.
Define MCP as an open protocol for LLM agents to connect to tools, data, and prompts.
Attribute MCP correctly: Anthropic, introduced late 2024.
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.