What is MCP (Model Context Protocol)? A PM guide
Introduced by Anthropic in November 2024, MCP (Model Context Protocol) is an open specification that defines how AI applications connect to external tools and data sources. An MCP server exposes capabilities, mainly tools, resources and prompts, and any compatible client (an AI assistant, an IDE, an agent) can use them without a custom integration. It is often compared to a universal port: one connector instead of one integration per pair of app and tool.
Definition written by Alex Dionisio, Senior Product Manager. Updated November 13, 2026.
Why it matters for a PM
MCP raises two product questions. As a user of AI tools, you decide which servers your team may connect, with what access, and how to vet them. As a product owner, you may consider exposing your own product as an MCP server so that assistants can act on it, which is a distribution and security decision as much as a technical one.
Example
A product team connects its AI assistant to the ticketing system through an MCP server in read-only mode. PMs can ask “which bugs tagged checkout were reported this week?” and get answers from live data, while creating or closing tickets stays out of reach until the team has reviewed the risks.
How does MCP work?
MCP defines three roles. The host is whichever AI application you work in, such as a desktop assistant or a code editor. A single host can be connected to many servers at once. For each server it connects to, the host creates a dedicated MCP client. The server is the program that gives access to a system: a ticketing tool, a file store, a database. Messages follow the JSON-RPC 2.0 format, and each side declares its capabilities so the other can adjust what it asks for.
A server exposes three kinds of building blocks. Tools are functions the application runs when the model decides to use them, such as creating a ticket. Resources are data the application can add to the context, such as the contents of a file. Prompts are reusable interaction templates, often offered to the user as commands. The specification also lets a server ask the user for extra information mid-task, a mechanism called elicitation.
MCP or API: what is the difference?
An API is designed for developers who write code around it. An MCP server is designed for a model that has to work out on its own what it can do: every tool comes with a name, a plain-language description and a parameter schema. The server usually sits on top of the existing API, but it chooses what to expose and how to present it.
That difference changes the design. An API may offer hundreds of endpoints; a good MCP server groups a handful of them around useful tasks, with precise descriptions and error messages the model can make sense of. Too many tools, or badly named ones, and the model picks the wrong one. Some teams also publish separate servers for reading and for writing, so access can be granted in stages.
How do you vet an MCP server before connecting it?
Apply the questions of a security review to this new party. Who publishes the server, and is it actively maintained? Does it run on the user's machine or remotely, and if remotely, which data passes through that service? What permissions does it request, and can you restrict it to read-only access? Are the tool descriptions readable and consistent with what the server claims to do?
The specification asks hosts to obtain explicit user consent before invoking a tool. Check how your host applies it: a confirmation on every write call is solid protection, a blanket approval granted once and for all is much weaker.
Finally, keep an inventory. Knowing which servers are connected, by whom and with which accounts makes it possible to revoke access quickly when a server is compromised or abandoned by its maintainer.
Key points
- MCP standardizes the connection between a client and a server; it does not decide what the model does with the tools.
- Servers run locally (stdio transport) or remotely over HTTP, with OAuth-based authorization for remote access.
- An MCP server often wraps an existing API, so it does not replace good API design.
- Starting read-only, with least privilege, limits the damage of a misbehaving server or model.
Common mistakes
- Installing community servers without checking who maintains them and what they can access.
- Assuming MCP handles authorization and data filtering for you.
- Overlooking tool poisoning: malicious instructions hidden in tool descriptions or results.
- Building an MCP server before knowing which user tasks it should serve.
Go further with Module
The courses and lessons that cover this concept:
Guides on this topic
- Claude Code for Product Managers
What a PM can do with Claude Code, the method and the guardrails, with a starter CLAUDE.md.
- Claude Skills vs MCP: which to use
What sets a skill apart from an MCP server, when to use one, the other or both, with a decision grid.
- MCP for Product Managers
Connect AI to your tools and scope an MCP server: tools, permissions, security, with a spec template.