What separates this from a template repository is what it puts in the directory: the downloaded SDK documentation and the rule files exist so the implementing agent is not working from memory of the protocol. The conversational state is a real design decision too — each answer is a separate call, which suits clients that cannot run a multi-step form.
An MCP server whose job is to create MCP server projects. Rather than a one-shot generator, it holds conversational state across tool calls — name, directory, language, description — and then builds a project directory that includes a PRD, an implementation guide, coding-standard rule files and downloaded SDK documentation, so the AI agent that writes the code has the context it needs.
- A step-by-step flow: `start_mcp_project`, then `set_project_name`, `set_project_directory`, `set_project_technology` (typescript or python) and `set_project_description`
- `add_project_documentation` — attach extra context before generation, for projects with an existing spec
- `setup_project_foundation` — create the directory structure and download the essential MCP documentation into it
- `generate_mcp_server` — produce the final context set, leaving `README.md`, `CLAUDE.md`, `IMPLEMENTATION.md`, `PRD.md`, rule files under `.windsurf/rules/` and SDK docs under `docs/external/`
- A tools-only design with no prompts, so it works in clients that support tools alone
Nothing — no account, no key. Node.js 18.0.0 or higher; `npm install` and `npm run build` from a clone, then point the client at `build/index.js` over stdio. Works on macOS, Linux and Windows.