Two things called "agent standards" get confused constantly: AGENTS.md, the context file that tells a coding agent how your repo works, and MCP, the protocol that gives an agent tools. They solve different problems and you'll usually want both. Here's the clean distinction — and why AGENTS.md is showing up in 60,000-plus repos.
What AGENTS.md is
AGENTS.md is a Markdown file at your repo root — a "README for agents." It carries build commands, coding conventions, testing rules and constraints an agent can't infer from the code alone. There are no required fields and no special syntax; it's plain Markdown a human can read too. Formalised as an open spec in 2025 and now governed under the Agentic AI Foundation, it's supported by Codex, Cursor, GitHub Copilot and more.
What MCP is
MCP is a protocol, not a document. It connects an agent to external tools and data — a database, GitHub, a search API — over a standard wire format so the agent can act, not just read instructions. Where AGENTS.md is static context loaded at the start of a task, MCP is a live connection the agent calls during the task, getting fresh results each time.
Instructions vs tools
The split is that simple: AGENTS.md tells an agent how to behave in your project; MCP servers give it things to do. One shapes conduct, the other grants capability. A well-run repo has an AGENTS.md that documents conventions and an MCP setup that wires in the tools — and they don't overlap, so there's no "which one" to pick between. You reach for both.
Using both
Put your conventions in AGENTS.md, then declare the servers the work needs. If your AGENTS.md says "run migrations with the Postgres tools," the Postgres server is what makes that instruction actionable. Instruction plus capability is the whole point: the file sets the rules, the servers do the work.
Going further
For the capability side, start with best MCP servers for developers and the Claude Code setup guide. See where Skills fit vs MCP, and what an MCP server actually is.