MCP Server Builder

The MCP "hello world" is trivial — the SDK hands it to you. The hard part is everything that makes a server safe to deploy: MCP tools are arbitrary code execution exposed to an LLM, tool descriptions are untrusted text the model reads, and remote servers are network services handling credentials. This skill builds the server and closes those gaps up front, instead of leaving them to a later audit.

It locks the key decisions first — TypeScript vs Python/FastMCP, stdio vs Streamable HTTP, which primitives, what the tools actually touch — then:

• Scaffolds the project on the official SDK (pinned), JSON-RPC 2.0 under the hood
• Defines each tool with a tight JSON Schema and a narrow, single-purpose scope
• Implements handlers defensively — least privilege, command/path/SSRF handled
• Wires the transport correctly (clean stdout for stdio; HTTPS + sessions for remote)
• Adds OAuth 2.1 + PKCE + resource indicators for remote, with no token passthrough
• Returns structured, sanitized errors that never leak internals or secrets
• Gates destructive actions behind confirmation

You get a runnable server skeleton with a fully implemented example tool, plus a Security Checklist Report showing how each footgun is handled and what you still need to wire (e.g., your OAuth provider). It defers to the current MCP spec and official SDK for exact APIs, and doesn't replace running the MCP Inspector and your own tests.

To audit a server or tool definitions you didn't build — schemas, descriptions, scopes, and auth checked for a SAFE / REVIEW / BLOCK verdict — pair it with the MCP Server & Tool-Definition Security Gate.