Model Context Protocol (MCP)

Date: 2026-03-31 Category: standards-research Status: Complete

Research Question

What is MCP, how does it differ from Agent Skills, and when should it be used instead of or alongside SKILL.md-based skills?

Background

The PRD references MCP without clearly distinguishing it from the skill system. Claude Desktop uses MCP servers for extension; Claude Code and Cursor use SKILL.md. This research clarifies the relationship and recommends when to use each.

Methodology

  • Reviewed the official MCP specification at modelcontextprotocol.io
  • Reviewed MCP tool and resource specifications
  • Reviewed Claude Desktop and Claude Code MCP configuration guides
  • Analyzed the architectural differences between MCP and Agent Skills

Key Findings

Finding 1: MCP Is an Open Standard for AI-Tool Integration

  • Description: Model Context Protocol (MCP) is an open standard created by Anthropic (released late 2024) that enables LLM applications to connect to external tools, databases, and APIs via a standardized JSON-RPC 2.0 protocol. As of 2026, it is the de facto standard for tool integration in AI agents.
  • Evidence: Official specification published at modelcontextprotocol.io
  • Confidence: High
  • Source: MCP Specification

Finding 2: MCP Exposes Three Capability Types

  • Description: MCP servers expose:
    1. Tools: Functions the model can call (e.g., search_web, query_database)
    2. Resources: Read-only structured data (e.g., file contents, DB records)
    3. Prompts: Reusable prompt templates for standardized tasks
  • Evidence: Specified in the MCP server capabilities spec
  • Confidence: High
  • Source: MCP Tools, MCP Resources

Finding 3: MCP Transport Options

  • Description: MCP servers communicate via two transport methods:
    • stdio: Local subprocess communication (zero network overhead, best for local tools)
    • SSE/HTTP: Server-Sent Events over HTTP (for remote/multi-user deployments)
  • Evidence: Specification defines both transport protocols
  • Confidence: High
  • Source: MCP Specification

Finding 4: Claude Desktop Uses MCP for Extension

  • Description: Claude Desktop (the standalone app) extends via MCP servers configured in claude_desktop_config.json:
    • macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
    • Linux: ~/.config/claude/claude_desktop_config.json
    • Windows: %APPDATA%\Claude\claude_desktop_config.json Claude Desktop does NOT have a native skills directory; MCP is the only extension mechanism.
  • Evidence: Official configuration documentation
  • Confidence: High
  • Source: MCP Installation Guide

Finding 5: MCP vs Agent Skills Are Complementary

  • Description: MCP and Agent Skills serve different purposes:
    • Agent Skills (SKILL.md): Static knowledge, instructions, workflows, documentation references. No runtime required.
    • MCP Servers: Dynamic tools, live data access, API integration. Requires a running server process. For the rhel-devops-skills-cli use case (providing documentation and workflow guidance), Agent Skills are the appropriate mechanism. MCP would be appropriate if the skills needed to execute live tools (e.g., running ansible-navigator, querying APIs).
  • Evidence: Skills are file-based; MCP requires server processes
  • Confidence: High
  • Source: Claude Code MCP, Agent Skills Spec

Finding 6: MCP Registry (Preview)

  • Description: An MCP Registry is in preview at modelcontextprotocol.io/registry, providing a centralized metadata repository for publicly accessible MCP servers. This could be relevant for future distribution of skills as MCP servers.
  • Evidence: Registry documented as preview feature
  • Confidence: Medium
  • Source: MCP Registry

Implications

Architectural Impact

  • The primary use case (providing documentation and workflow guidance) is best served by Agent Skills (SKILL.md), not MCP servers
  • MCP support for Claude Desktop should be considered a future enhancement, not a primary target
  • If the project later needs live tool execution (e.g., running validation scripts, fetching live repo data), MCP servers would be the right approach

Technology Choices

  • Primary: Agent Skills (SKILL.md) for documentation and instructions
  • Future: MCP servers (Python with mcp SDK) for live tool integration
  • Config format: JSON for MCP server registration in claude_desktop_config.json

Risk Assessment

  • Low risk: Choosing Agent Skills for initial release (simpler, no runtime dependency)
  • Medium risk: Deferring MCP support excludes Claude Desktop users initially
  • Low risk: MCP can be added later without breaking existing skills

Recommendations

  1. Use Agent Skills (SKILL.md) as the primary delivery mechanism
  2. Defer MCP server packaging to a future release
  3. If Claude Desktop support is critical, create a lightweight MCP server that exposes documentation as Resources
  4. Document the MCP vs Skills distinction in user-facing docs
  5. Monitor the MCP Registry for potential distribution channel
  • ADR-002: Target Claude Code + Cursor
  • ADR-003: Documentation Embedding Strategy

References