Short answer: an MCP server is a small program that exposes tools and data to an AI assistant through a standard protocol, the Model Context Protocol. For AWS engineers, it is a controlled way to let a coding agent read docs, query resources or run approved actions, with permissions you define.
How MCP works
| Piece | Role |
|---|---|
| Host / client | The AI application, for example Claude Code or an IDE assistant |
| Server | Your program that exposes tools (actions), resources (data) and prompts |
| Transport | Local process over stdio, or remote over HTTP |
| Credentials | Whatever the server process uses, for example an AWS profile or role |
When an MCP server is worth it
- The agent needs live data it cannot guess: resource inventory, logs, costs, current docs.
- You want repeatable, reviewed actions instead of the agent improvising CLI commands.
- Several people or agents need the same integration with the same guardrails.
If the agent only needs to run a few CLI commands you already trust, a well-scoped AWS profile may be enough.
Add a server to Claude Code
shell
# Add a local (stdio) MCP server; everything after -- is the command that starts it
claude mcp add my-aws-docs -- uvx my-aws-docs-server
# List configured servers
claude mcp listAWS publishes open-source MCP servers for its documentation and several services. Check the project page for the current list and the exact install command.
An MCP server runs with its own credentials. Give it a read-only profile first, and add write actions one by one, with an explicit allow list. The agent can call any tool the server exposes.
A safe starting setup
- A dedicated IAM role for the server, read-only, in a non-production account.
- CloudTrail on, so every call made through the server is auditable.
- Tools that change resources require confirmation in the client.
- Review the server code or use a source you trust: it runs on your machine.