mcp-server-designlisted

Design an MCP server's tool surface — equilibrated recursivity (flat schema, deep data), 1 tool vs N tool decomposition, output contract, error-code discipline, capability negotiation, security checklist, tool naming, and the Claude-Optimal validation. Use when the user says "design an MCP server", "decompose MCP tools", "MCP schema design", "MCP tool naming", "MCP error codes", "MCP security checklist".
Git-Fg/taches-principled · ★ 0 · AI & Automation · score 73

Install: claude install-skill Git-Fg/taches-principled

# mcp-server-design The design layer for MCP servers: how to think about tool decomposition, output contracts, error codes, security, and naming. For the implementation in Rust (rmcp + schemars), see `mcp-server-implement`. For the JSON Schema authoring details, see `mcp-tool-surface`. ## §1. When this skill fires **Use this skill when the user says any of:** - "Design the tool surface for my MCP server" - "How many tools should I expose — 1 mega-tool or N small ones" - "MCP output contract" - "MCP error code discipline" - "MCP capability negotiation" - "MCP security checklist" - "What makes a Claude-Optimal MCP server" - "Wrap a CLI as an MCP server" → this skill + `mcp-server-implement` together; the `claude-cli` skill is a worked example ## CONTRAST - NOT for: writing a JSON Schema for a tool — use mcp-tool-surface - NOT for: implementing the server in code (rmcp, schemars, transport) — use mcp-server-implement - NOT for: building a non-MCP CLI tool — use subagent-orchestration or shell tooling - This skill is the design layer; the implementation layer is mcp-server-implement ## §2. Reference index The mechanism content lives in references/. Read the right one before committing to a design decision. The hub itself is a router — it points you at the right reference, the references carry the mechanism. You MUST read `references/design-decisions.md` BEFORE committing to a tool decomposition or output contract. It teaches the equilibrated-recursivity thesis (flat schem