Standardize MCP server config across AI agents

[codeaholicguy]

Summary

• Define MCP servers once in .ai-devkit.json (or YAML templates), generate agent-specific configs via ai-devkit install
• Support Claude Code (.mcp.json) and Codex (.codex/config.toml), extensible to more agents via mcpConfigPath in environment definitions
• Support stdio, http (streamable-http), and sse (deprecated) transport types
• Safe merge with existing configs, interactive prompt for conflicts, CI-safe with --overwrite flag

How it works

1. Define servers in a template or .ai-devkit.json, example:

mcpServers:
  memory:
    transport: stdio
    command: npx
    args: ["-y", "@ai-devkit/memory"]
  notion:
    transport: http
    url: https://mcp.notion.com/mcp

2. ai-devkit init -t template.yaml → persists mcpServers to .ai-devkit.json
3. ai-devkit install → generates .mcp.json and .codex/config.toml from config
4. Existing agent configs are preserved, only managed servers are added/updated

Key design decisions

• Single source of truth: mcpServers lives in .ai-devkit.json
• Abstract base class: BaseMcpGenerator handles shared plan/apply diff-merge logic; subclasses only provide format-specific I/O
• CI-safe: non-interactive mode skips conflicts by default; --overwrite forces overwrite; detected via shared isInteractiveTerminal() util
• mcpConfigPath on EnvironmentDefinition: add the path to enable MCP for a new agent