DR-008: MCP Server Architecture for AI Integration

Status

[x] Accepted
[ ] Proposed
[ ] Rejected
[ ] Deprecated
[ ] Superseded

Date: 2024-12-15

Context

Claude Code and other AI systems need programmatic access to repository operations (modules, dependencies, files, builds, tests) through structured interfaces.

Problem: How to bridge AI agents and repository operations?

Decision

Model Context Protocol (MCP) server exposing EAC commands as AI-accessible tools via stdio protocol.

Architecture:

Stdio-based MCP server (go/eac/mcp/commands)
Dynamic tool generation from command definitions
Tool naming: mcp__commands__<command-name>
Output formats: YAML/JSON for machine consumption

Dual Execution Model:

CLI mode: Human-readable tables and summaries (r2r eac show modules)
MCP mode: Structured data for AI parsing (mcp__commands__get-modules)

Claude Code Integration (.claude/):

6 specialized agents (go-architect, go-debugger, go-test-engineer, go-cli-ux, go-security-release, go-workflow-engineer)
3 skills (go-cli-feature, go-cli-release-check, go-cli-refactor-safe)
8 slash commands (/go:plan, /go:implement, /go:test, /go:review, /go:debug, /go:release, /go:cli-docs, /go:session-end)

MCP Configuration (.mcp.json):

{
  "mcpServers": {
    "commands": {
      "type": "stdio",
      "command": "go",
      "args": ["run", "./go/eac/mcp/commands/main.go"]
    }
  }
}

Consequences

Positive: AI-native workflows, reusable integration, structured data exchange, unified command interface, context-aware operations

Negative: MCP server maintenance, dual output complexity, stdio protocol limitations

Alternatives Considered

REST API: Rejected - overkill for local tools, server overhead, authentication complexity
Language Server Protocol (LSP): Rejected - focused on code editing, not repository operations
Custom JSON-RPC: Rejected - MCP is emerging standard, better ecosystem support

Tutorials | How-to Guides | Explanation | Reference

You are here: Reference — information-oriented technical descriptions of the system.