MCP Server Project Generator

Trigger

• •Use when given MCP folders or file lists containing mcp_init.md and asked to scaffold runnable MCP servers/connectors from those specs.

Inputs

• •One or more MCP units; each must include mcp_init.md.
• •Parse each spec for: MCP type (connector|server), name (external_api_name|domain_name), tools (name, purpose, request/response fields, data_source upstream|db|mixed, error cases), external integration (API base/auth/headers/query/body, DB intent), model requirements (DAO/DTO/service/repository/connection).
• •If data is missing, do not ask questions: choose reasonable defaults and mark “Assumption:” in README + TODO comments.

Outputs

• •For every MCP, emit full project tree and full code per file (path + content). Project must be runnable (scaffold level). Unimplemented logic stays as TODO with interfaces shaped.
• •Absolutely no pytest/테스트 files.

Global Conventions

• •FastMCP boilerplate (must keep order/shape):
  from fastmcp import FastMCP
mcp = FastMCP("<mcp_name>")
@mcp.tool (or @mcp.tools only if spec forces; note it in README)
  tool def/async def …
  if __name__ == "__main__": mcp.run()
• •Env vars (NAME = upper snake of external_api_name/domain_name):
  LOG_LEVEL (default INFO); DB_DSN (server only); MCP_<NAME>_BASE_URL; MCP_<NAME>_API_KEY; MCP_<NAME>_TIMEOUT_SEC (default 20); MCP_<NAME>_RETRY (default 2).
• •Types mapping: string→str, integer/int→int, float/double/number→float, boolean→bool, datetime→datetime.datetime (ISO8601 parse), date→datetime.date, decimal→decimal.Decimal, object/dict→dict[str, Any], array/list→list[T or Any]. Optional/nullable/? → Optional[T]=None. Enums → Literal[...] or Enum (Literal preferred). Unknown → Any with “Assumption:” notes.

Required Structure (unless spec overrides)

<root>/
  mcp/
    common/
      settings.py      # Pydantic Settings loading .env
      logging.py       # std logging with request_id helper
      errors.py        # ValidationError/UpstreamError/DatabaseError/InternalError + payload mapper
      http_client.py   # httpx client factory with timeout/retry
      db.py            # SQLAlchemy engine + SessionLocal (tool-scope)
      types.py         # Meta, ErrorPayload, EnvelopeResponse
    connectors/<external_api_name>/
      mcp_connector_<external_api_name>.py
      tools/__init__.py, <tool_group>.py
      connection/__init__.py, client.py, schemas.py
      models/__init__.py, dto.py
    servers/<domain_name>/
      mcp_server_<domain_name>.py
      tools/__init__.py, <tool_group>.py
      service/__init__.py, interface.py (interface only, TODO impl)
      repository/__init__.py, repo.py (SQLAlchemy)
      models/__init__.py, dao.py (ORM), dto.py (Pydantic)
  README.md
  pyproject.toml or requirements.txt

Behavior Rules

• •Connector: No service/repository. tools → connection.client. All upstream→internal DTO transforms live only in connection/client.py; tools never touch raw upstream. Entry file named mcp_connector_<external_api_name>.py.
• •Server: tools → service (interface only) → repository → db. Entry file mcp_server_<domain_name>.py. DAO uses declarative base. Repository gets session injected. DB session scope is per tool:
  with SessionLocal() as session: repo = ...; ...
• •Tool design: exactly one Request DTO and one Response DTO. Docstring must include purpose, input fields, output fields, error codes, brief example. meta.source = upstream (connector), db (server DB-only), mixed (server API+DB).
• •Envelope response (never raise to caller):
  code

  {"data": <ResponseDTO|dict>,
   "meta": {"request_id": "...", "source": "upstream|db|mixed", "generated_at": "ISO8601", "version": "v1"},
   "error": null | {"code": "VALIDATION_ERROR|UPSTREAM_ERROR|DATABASE_ERROR|INTERNAL_ERROR", "message": "...", "detail": {...}}}
  

Workflow

1. •Discover all mcp_init.md from provided folders/files. If a referenced path is missing, note as TODO and proceed with available ones.
2. •For each spec, extract fields (type/name/tools/etc). Infer missing data with defaults; log “Assumption:” in README + code TODO.
3. •Decide MCP type → choose connector/server layout and env var NAME.
4. •Generate common layer once per project.
5. •Per MCP, create entrypoint, tools, DTOs, connection/client (connector) or service/repository/models (server). Keep interfaces complete, logic TODO where spec is silent.
6. •Ensure every tool returns Envelope and handles Validation/Upstream/Database/Internal errors into payload.
7. •Write README.md covering MCP overview, tools, env vars, run instructions, Assumptions, TODOs.
8. •Output: list every created file with path + full content. No pytest files.

Notes

• •Keep instructions concise in outputs; avoid extra assets unless required.
• •Respect any spec-enforced import paths/decorators, but preserve FastMCP boilerplate shape.
• •Never prompt the user for clarifications. Use explicit TODO + Assumption markers instead.