mcp-export-design

Designs dual-era MCP server export surfaces projected from a typed internal tool IR.

Updated Aug 2, 2026
One-click install
npx skills add https://github.com/dperussina/function2agent --skill mcp-export-design-dperussina
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: mcp-export-design
Source: https://github.com/dperussina/function2agent/tree/main/.cursor/skills/mcp-export-design
Command: npx skills add https://github.com/dperussina/function2agent --skill mcp-export-design-dperussina

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? Teams generating MCP servers from a codebase risk shipping servers that break against every shipping client, lose safety-critical metadata (read_only, egress, idempotence) that JSON Schema has no slot for, and wrongly assume MCP provides progressive disclosure, tool search, or code-mode execution. ## Core Features & Use Cases - Architecture decision guidance: Keeps a typed internal IR as the single source of truth and projects MCP as an export adapter, never as the internal calling convention. - Protocol churn planning: Covers the four breaking MCP revisions (2024-11-05 through 2026-07-28) and mandates dual-era generated servers that serve legacy initialize and modern per-request _meta on one endpoint. - Emission checklist: Provides a concrete checklist for generated servers covering MRTR input_required gating, handle-passing for cross-call state, deterministic tools/list ordering, and JSON Schema 2020-12 input schemas. - Use Case: When designing an MCP server that exposes a generated tool catalog to Claude, ChatGPT, Cursor, and Gemini, use this Skill to decide the wire revision, transport, and catalog size before writing any server code. ## Quick Start Ask the agent to design the MCP export surface for your generated tool catalog and verify the emitted server against the dual-era compatibility checklist.

Frequently Asked Questions about mcp-export-design

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
Should MCP be the internal tool representation for an agent?▼

No. MCP schemas have no slot for safety metadata like read_only, egress, idempotence, or auth scope, so making MCP canonical leaves the safety story unrepresentable. Keep a typed internal IR as the source of truth and project MCP from it.

How do I make a generated MCP server work with legacy clients?▼

Generate dual-era servers that serve legacy initialize and modern per-request _meta concurrently on the same endpoint, which the spec explicitly permits. A modern-only server fails against every MCP client shipping today with no downward retry path.

Does the MCP protocol support tool search or progressive disclosure?▼

No. defer_loading, tool search, and allowed_callers are proprietary Anthropic and OpenAI API constructs, not MCP protocol features. A generated server's only real levers are exposing fewer tools and naming them well at generation time.

Can MCP tools be called programmatically in code mode?▼

Tools provided through Anthropic's MCP connector cannot be called programmatically. Code mode works only when you call tools you define directly in the Messages API or when you own the client and sandbox calling the MCP server.

How stable is the MCP wire format across revisions?▼

The protocol shipped four breaking revisions in roughly 20 months, and the 2026-07-28 revision removed sessions, initialize, ping, and SSE resumability entirely. Plan generated servers for continued churn rather than assuming Linux Foundation stewardship means stability.

How should destructive tools be gated in a generated MCP server?▼

Gate destructive tools via MRTR input_required results rather than a model-callable confirm tool, and state destructiveness explicitly in description prose since no enforceable schema slot exists. Keep the effects metadata in the internal IR.