Claude Code Plugins

Community-maintained marketplace

Feedback

This skill should be used when the user asks to "create an MCP server", "build MCP tools", "define MCP prompts", "register MCP resources", "implement Model Context Protocol", or mentions the mcp gem, MCP::Server, MCP::Tool, JSON-RPC transport, stdio transport, or streamable HTTP transport. Should also be used when editing MCP server files, working with tool/prompt/resource definitions, or discussing LLM tool integrations in Ruby.

Install Skill

1Download skill
2Enable skills in Claude

Open claude.ai/settings/capabilities and find the "Skills" section

3Upload to Claude

Click "Upload skill" and select the downloaded ZIP file

Note: Please verify skill by going through its instructions before using it.

SKILL.md

name MCP Server (Ruby)
version 1.0.0
description This skill should be used when the user asks to "create an MCP server", "build MCP tools", "define MCP prompts", "register MCP resources", "implement Model Context Protocol", or mentions the mcp gem, MCP::Server, MCP::Tool, JSON-RPC transport, stdio transport, or streamable HTTP transport. Should also be used when editing MCP server files, working with tool/prompt/resource definitions, or discussing LLM tool integrations in Ruby.

MCP Ruby SDK - Server Development Guide

Build Model Context Protocol servers in Ruby using the official mcp gem (maintained by Anthropic and Shopify).

Design Philosophy

Information Provider, Not Analyzer

MCP servers provide structured data; LLMs do the reasoning. Return comprehensive frameworks and raw information—let the client perform analysis and context-dependent decisions.

"The MCP server's job is to be the world's best research assistant, not a competing analyst." — Matt Adams

Context Preservation

Agents have limited context windows. Every byte returned that wasn't requested is a byte that could have held useful context. Treat context preservation as a first-class design constraint.

Principles:

  • Never return data that wasn't explicitly requested
  • Mutations are quiet—return confirmations, not data dumps
  • Explicit over implicit—associations only when asked
  • Filter large datasets before returning (10,000 rows → 5 relevant rows)

Domain-Aligned Vocabulary

Tools should speak the language of your domain, not database/CRUD terminology. Agents are collaborators in your domain process, not database clients.

Example: A visual novel asset server uses create_image, make_sprite, place_character, explore_variations, compare_images—not generate, remove_background, composite, batch_generate, get_diff.

Tool Budget Management

Too many tools overwhelm agents and increase costs. Design toolsets around clear use cases, not API endpoint mirrors.

  • Group related functionality intelligently
  • Use lazy loading for large tool sets (150K tokens → 2K via on-demand discovery)
  • Tool names ≤64 characters, descriptions narrow and unambiguous

Security: The Lethal Trifecta

Three capabilities that, when combined, create vulnerabilities (Simon Willison):

  1. Access to private data
  2. Exposure to untrusted content
  3. External communication capabilities

Required: Explicit user consent before tool invocation, clear UI showing exposed tools, alerts when tool descriptions change.

Domain Components

Component Purpose Reference
Tools Define callable functions with input/output schemas `references/tools.md`
Prompts Template-based message generators `references/prompts.md`
Resources Static and dynamic file/data registration `references/resources.md`
Server Core server initialization and configuration `references/server.md`
Transport STDIO and HTTP transport options `references/transport.md`
Gotchas Tricky behaviors and error handling `references/gotchas.md`

Key Concepts

Concept Purpose
MCP::Tool Base class for defining callable tools
MCP::Prompt Base class for prompt templates
MCP::Resource Static resource registration
MCP::ResourceTemplate Dynamic URI-based resources
server_context Request-scoped data passed to handlers
MCP::Tool::Response Structured tool return value

Tool Definition Patterns

Pattern Use Case
Class-based (< MCP::Tool) Reusable tools with complex logic
Block-based (MCP::Tool.define) Inline, simple tools
Dynamic (server.define_tool) Runtime tool registration

Transport Decision Tree

What environment?
├── CLI tool / Local server
│   └── Use STDIO transport
└── Web server / Production
    └── Need sessions and notifications?
        ├── YES → Use Streamable HTTP (stateful)
        └── NO → Use Streamable HTTP (stateless)

Quick Comparison

Transport Sessions Notifications Use For
STDIO N/A Yes CLI tools, local dev
HTTP (stateful) Yes Yes Web apps, long-lived connections
HTTP (stateless) No No Simple request/response APIs

Protocol Version Features

Feature Minimum Version
description 2025-11-25
instructions 2025-03-26
annotations 2025-03-26
output_schema 2025-03-26

Best Practices

Do

  • Use tool_name for namespaced classes to avoid conflicts
  • Use additionalProperties: false for strict schema validation
  • Use mutex for shared state in HTTP transport (thread safety)
  • Return error responses for business errors (Response.new([...], error: true))
  • Check protocol version before using newer features
  • Use server_context for request-scoped data (user_id, env)

Don't

  • Don't use $ref in schemas (raises ArgumentError, inline only)
  • Don't assume extra args are rejected (additionalProperties defaults to allowing extras)
  • Don't use rpc. prefix (reserved for protocol methods)
  • Don't send notifications in stateless mode (raises RuntimeError)
  • Don't rely on validation order (required args checked before JSON Schema)

Anti-Patterns Quick List

Anti-Pattern Solution
Missing additionalProperties: false Add to schema for strict validation
Using $ref in schemas Inline all definitions
Notifications in stateless mode Use stateful transport or skip notifications
Hardcoded server_context Pass dynamically based on request
Ignoring protocol version Check version before using gated features
Blocking in tool handlers Use async patterns for long operations

Key Points

  1. Validation is multi-layered - Required args checked first, then JSON Schema validation
  2. Notifications are fire-and-forget - Errors reported but don't propagate
  3. Protocol version matters - Features are gated by version
  4. Server context is opt-in - Detected from method signature (must include server_context: parameter)
  5. Schemas are immutable - Validated at class load time, not runtime

Additional Resources

Reference Files

For detailed DSL syntax by domain:

  • references/tools.md - Tool definition, responses, schemas, annotations
  • references/prompts.md - Prompt definition, arguments, content types
  • references/resources.md - Resource registration, templates, read handlers
  • references/server.md - Server initialization, configuration, custom methods
  • references/transport.md - Transport config, protocol methods, sessions
  • references/gotchas.md - Tricky behaviors, error handling, edge cases

Example Files

Working examples in examples/:

  • examples/stdio_server.rb - Complete STDIO server with tools, prompts, resources
  • examples/http_server.rb - HTTP server with Rack and logging
  • examples/rails_integration.rb - Rails controller, routes, and initializer
  • examples/file_manager_tool.rb - Sandboxed file operations with security patterns
  • examples/dynamic_tools.rb - Runtime tool registration with notifications
  • examples/http_client.rb - HTTP client connecting to MCP server
  • examples/streaming_client.rb - SSE streaming client for real-time notifications

External Links