Skip to content

Tool Interfaces and MCP — Junior

At junior level, focus on this question:

Can you explain what MCP actually is — a protocol, not a product — and trace one tool call from the model discovering it exists to a result coming back?

What MCP is

The Model Context Protocol (MCP) is an open, standardized protocol for connecting an AI application to external tools, data sources, and prompts, through a client-server architecture. (Model Context Protocol specification)

  • An MCP server exposes capabilities in three kinds: tools (actions the model can invoke, e.g., run_query), resources (data the client can read, e.g., a file or a dataset schema), and prompts (reusable prompt templates the server provides).
  • An MCP client (the agent framework) connects to one or more servers, discovers what they expose, and makes those capabilities available to the model.
  • Transport is the connection mechanism — commonly stdio (the server runs as a local subprocess, communicating over standard input/output) or HTTP-based (the server runs remotely, reachable over a network).
  • The protocol standardizes how an agent discovers and calls tools/resources — before MCP, every agent framework and every tool integration used its own bespoke interface; MCP is a common contract so a server built once can be used by any compliant client.

One round trip, traced

sequenceDiagram participant A as Agent (MCP client) participant S as MCP server (BigQuery) participant M as Model A->>S: list available tools (discovery) S-->>A: tool schemas (e.g., run_query, list_datasets) A->>M: messages + discovered tool schemas M-->>A: tool_use: run_query(sql="SELECT ...") A->>S: execute run_query S-->>A: result rows A->>M: tool result appended M-->>A: final answer
  1. Discovery — at connection time, the client asks the server what it can do; the server returns its tool/resource/prompt schemas.
  2. Exposure — the client includes those schemas in the model's available tools, the same way any function-calling schema is exposed to a model as a JSON tool definition.
  3. Invocation — the model emits a structured call naming a tool and its arguments; the client routes this to the correct MCP server.
  4. Execution and result — the server runs the actual operation (e.g., a BigQuery query) and returns a structured result to the client, which appends it to the conversation for the model to reason over.

Why this matters for context, specifically

Every tool an MCP server exposes is discovered once per session and then occupies context on every subsequent turn as part of the available tool schemas — connecting to a server with 30 tools costs tokens on every turn even if the agent only ever calls 2 of them. This is the same tax as any other tool schema (see Context Fundamentals); MCP doesn't make it free, it standardizes how the tool got there.

Comprehension check

  • What are the three kinds of things an MCP server can expose, and give one example of each.
  • What problem does MCP's standardization actually solve, compared to a bespoke per-framework tool integration?
  • Trace the four steps of one MCP tool call round trip in your own words.
  • Why does connecting to an MCP server with 30 tools cost tokens even on a turn where none of them are called?