MCP Explained Without the Hype

The model asks. The client brokers. The server exposes capabilities.

MCP is an open protocol for connecting AI hosts to servers that expose capabilities. The point is not magic. The point is a shared contract, so builders do not need custom glue code for every app, tool, and data source.

• Host: the AI app or agent experience where the user works.
• Client: the protocol-speaking connector inside that host.
• Server: the service that exposes capabilities over MCP.
• Capability: a tool, resource, or prompt the assistant can discover and use.

Tools, resources, and prompts are the core pieces.

A useful MCP setup is not magic. It is a small menu of capabilities the assistant can call when the workflow needs them.

• If the model needs to do something, think tool.
• If the model needs to read something, think resource.
• If the team needs a repeatable instruction pattern, think prompt.

Use MCP when the workflow repeats.

If you only need one answer, chat may be enough. If the assistant repeatedly needs the same tool, file, API, or system of record, MCP starts to make sense.

• Good fit: repeated work with clear inputs and outputs.
• Good fit: read-only access to context the assistant needs often.
• Good fit: team workflows where the same tool connection can be reused.
• Bad fit: one-off tasks, unclear data boundaries, or sensitive writes without approval.

Start with one safe integration.

The first win is not a giant agent stack. The first win is a narrow capability that removes repeated copy-paste and still gives humans control.

• Name one workflow you repeat every week.
• Write the outcome in one sentence.
• List the minimum context the assistant needs.
• Define the exact jobs the server should support.
• Start with the minimum set of tools and resources.
• Write tight input schemas and predictable outputs.
• Choose read-only access first.
• Decide transport and authentication before rollout.
• Add a human review point before any write action.
• Log what the tool did and what it touched.
• Test with an inspector or low-risk workflow before broader use.

MCP makes permissions a product decision.

Once an assistant can use tools, security stops being an afterthought. The safest MCP workflow is scoped, inspectable, and easy to turn off.

• Use authentication for every sensitive endpoint.
• Authorize per tool and per resource, not only per server.
• Use least privilege credentials.
• Add explicit allowlists for dangerous actions.
• Keep secrets out of prompts and generated content.
• Prefer read-only until the workflow proves value.
• Require approval before writes, deletes, payments, or messages.
• Use rate limits, timeouts, and clear failure behavior.
• Keep logs so a human can understand what happened.

Keep these terms straight

MCP

Model Context Protocol, a standard for connecting AI apps to external capabilities.

Host

The AI app or agent environment where the user works.

Client

The protocol layer that connects the host to MCP servers.

Server

The service that exposes tools, resources, or prompts.

Tool

A callable action the assistant can use.

Resource

Context the assistant can read.

Prompt

A reusable workflow instruction.

Transport

How the client and server communicate, often local stdio or remote HTTP.

Scope

The boundary around what the assistant can access or do.

Repos worth studying