MCP

MCP, the Model Context Protocol, is an open standard for connecting a harness to outside capabilities. An MCP server is a small program that advertises tools (and sometimes resources and prompt templates) over a simple JSON interface; the harness lists those tools next to its built-in ones, and the model calls them the same way. One server for GitHub, one for Postgres, one for your browser, and any MCP-aware client can use all three without custom glue.

The pitch is portability: write the integration once, use it from Claude Code, Cursor, Codex, Claude Cowork and the rest. The catch is context. Every connected server adds its tool definitions to every request, and popular servers ship dozens of tools. Five servers can quietly eat a large slice of your context window before you type a word, and a crowded tool list makes the agent worse at choosing. A slow or flaky server also stalls the whole loop while the tool call waits.

Connect servers per project, not globally, and only the ones the current work needs. Prefer servers with a few well-described tools over kitchen-sink ones. For anything that touches production data, check what permissions the server holds; from the model's point of view an MCP tool is just another action it is allowed to take, and prompt injection through a server's results is a real path.