MCP: give AI assistants real tools

The Model Context Protocol is a standard plug shape between assistants and tools: one MCP server exposes files, a browser, a database or an API, and any MCP-speaking assistant can use it without custom glue. Before MCP every assistant needed its own plugin format; now the server is written once.

Start with two servers: filesystem (scoped to the project directory, never the whole disk) and fetch (for reading URLs). Add database and browser servers only when a task actually needs them — every extra tool is extra ways for the agent to surprise you.

Writing your own server

A minimal server wraps one thing your team does daily: run the test suite, query the staging database, file a ticket. Expose three to five operations with strict schemas rather than fifty thin wrappers — agents choose well from short menus and flail in long ones.

Version the server with the workflow it serves, log every call with arguments, and treat a confusing error message as a bug: the agent reads errors the way users read them, literally.

Security boundaries

Every MCP server is remote code execution with manners. Scope filesystem roots narrowly, require approval for writes and external calls, and never hand a server your production credentials when a read-only replica answers the same questions. Review third-party servers like dependencies: pin versions and read the diff on update.