There was a thread the other day about how there’s too many conflicting “standards” for where to add agent instructions to a repo (.cursorrules, AGENTS.md, etc).

I’ve got a new project which uses MCP to expose your project specific prompts (`testing-standards.md`, `design-guide.md`), and tools (`test`, `lint`, `format`, `build`) via MCP. All it takes is one YAML file to expose a custom MCP server, designed for your project.

https://github.com/scosman/hooks_mcp

- Discovery: the MCP server describes it’s functionality to your agents, both tools and prompts.

- Easy Setup: one YAML file to describe the tools (`npm check_format`) or prompts (`agents/front-end-design-guide.md`).

- Security and Speed: let your agents run some commands without approval (format, lint, etc), without granting wide open terminal access. MCP supports parallel tools for fewer chat turns.

- Portable Standard(??): MCP seems to be supported by all major agent tools. Some don’t support the ‘prompts’ feature of MCP, but for those we simply expose prompts via a `get_prompt` tool as a fallback. Also, xkcd N+1 standards joke.