MCP and code comprehension: keep understanding the code your agent changes

What is MCP, and what does it have to do with code comprehension?

MCP is a protocol that lets an AI application, such as a coding agent, call tools and read data that a separate program, called an MCP server, provides. The MCP documentation defines the protocol in one sentence:

Code comprehension enters in two ways. An MCP server can give the agent more information about your code, which helps the agent understand the code. Or an MCP server can ask you about the code, which checks whether you understand it. Most servers described as code-comprehension servers do the first.

Can an MCP server help a coding agent understand a codebase?

Yes, by giving the agent tools to find and read code. An MCP server exposes tools, and the agent decides when to call each tool. The MCP specification says tools are model-controlled, meaning that the language model can discover and invoke tools automatically based on its contextual understanding and the user's prompts.

For code, the tools are things like searching the repository, looking up where a function is defined, or fetching a library's current documentation. The agent calls a tool, reads the result, and uses it to write or change code. Claude Code's documentation describes the effect this way: MCP servers give Claude Code access to your tools, databases, and APIs.

In Claude Code, you add a server with claude mcp add. A remote server takes the form claude mcp add --transport http <name> <url>. A server added with --scope project is written to a .mcp.json file at the project root, which you can commit so that everyone on the team gets the same tools.

Does a better-informed agent mean you understand the code?

No. Tool results go to the model, not to you, so a code-search server gives the agent more information without teaching you anything. You still read only what the agent chooses to show you.

There is evidence that the gap between what the AI knows and what you know is real. In an Anthropic trial, 52 developers learned a new Python library. The developers who had an AI chat assistant scored about 15 points out of 100 lower on a quiz afterwards than developers who used documentation and web search. The largest gap was on debugging questions, about 21 points (Shen & Tamkin, 2026).

The trial has limits: one library, one 35-minute task, a quiz in the same sitting, and a paper not yet peer-reviewed. It also used a chat assistant, not an agent. The authors note that agentic tools need less participation from the developer than chat, so they expect developers who use agents to leave at least as much of the thinking to the AI.

How can you use MCP to check your own understanding?

Use a server whose tool returns a question for you instead of data for the agent. The agent calls the tool when it finishes a task, prints the question, and you answer from memory. For this to measure you and not the agent, the server has to meet four conditions:

• The server never sends the correct answer to the agent. An answer in the tool result is an answer the agent can give for you.
• The questions cover the topics of the task: the library, the concept or the kind of file that changed. You answer about the ideas, not about line 40 of a file you can scroll back to.
• You answer before you see the explanation, and then you see whether you were right and why.
• The server limits how often questions appear. A question after every small task interrupts your work until you turn the server off.

Many clients ask for your approval before a tool call, because the MCP specification recommends that a person can refuse any tool call. Check your client's permission settings for a way to allow the question tool without an approval prompt before every question.

Does answering questions from memory help you keep what you learned?

Yes. Answering questions from memory is called retrieval practice, and in the studies that compare the two, it produces better recall than reading the material again. In the reference experiment, students who practiced recalling a prose passage remembered 61% of it a week later, and students who reread it four times remembered 40% (Roediger & Karpicke, 2006).

A 2021 meta-analysis of 222 classroom studies, covering 48,478 students, found that quizzing raised achievement by about half a standard deviation (Yang et al., 2021).

The evidence is weakest where engineers need it most. There are about six studies on working professionals, their pooled result has a confidence interval that crosses zero, and there is no randomized trial on programming.

Which approach to code comprehension should you use?

Use a code-context server for the agent and questions for yourself, because the two help different people. This table compares a code-context server, asking the agent to explain its work, and a question server.

In the Anthropic trial, the developers who asked the assistant conceptual questions or asked for explanations along with the code scored 65-86% on the quiz. The developers who delegated the task or pasted errors back until the errors went away scored 24-39%. Each group had two to seven people, so treat the split as a pattern, not proof (Shen & Tamkin, 2026).