DevToolkit MCP Server

A production-quality Model Context Protocol (MCP) server that gives AI assistants like Claude 22 powerful developer tools across 5 domains.

Built with TypeScript, strict types, Zod validation, and clean modular architecture.

Features

📁 File Tools (6)

🔧 Git Tools (6)

🔍 Code Analysis Tools (6)

🌐 HTTP Tools (4)

💻 System Tools (6)

Installation

git clone https://github.com/your-username/devtoolkit-mcp
cd devtoolkit-mcp
npm install
npm run build

Usage with Claude Desktop

Add to your Claude Desktop config (claude_desktop_config.json):

{
  "mcpServers": {
    "devtoolkit": {
      "command": "node",
      "args": ["/absolute/path/to/devtoolkit-mcp/dist/index.js"]
    }
  }
}

Usage with Cursor / VS Code

Same pattern — point the MCP server config to dist/index.js.

Development

npm run dev    # Run with tsx (no build step)
npm run build  # Compile TypeScript
npm run lint   # ESLint

Advanced MCP Protocol Features

Beyond tools, this server implements the full 2025-06-18 MCP specification:

📂 Resources

Files are exposed as browsable MCP resources — clients (e.g. Claude Desktop) can list, read, and subscribe to changes.

• Paginated listing with cursor-based pagination (50 files/page)

• Resource templates: file:///{path} for dynamic file access

• Subscriptions: clients receive notifications/resources/updated when a file changes after a write/delete

• List-changed notifications: resource list updates after any file mutation

• Annotations: every resource includes audience, priority, and lastModified

📝 Logging

Structured log notifications sent to the client for every tool call:

• 8 RFC 5424 severity levels: debug → emergency

• Client can set minimum log level via logging/setLevel

• Logs include tool name, args preview, status, and errors

• Sensitive values never logged

🔍 Completion

Autocomplete for file paths in resource URI templates:

• Triggered when client requests completion/complete on a ref/resource

• Returns matching file paths up to 20 suggestions with hasMore flag

🗣️ Elicitation

Servers request structured user input before destructive operations:

• write_file: if the target file exists, asks user to confirm overwrite (with checkbox UI)

• delete_file: always asks user to confirm + optionally provide a reason

• Falls back gracefully if client doesn't support elicitation (proceeds without prompt)

🤖 Sampling

The server can ask the client's LLM to generate completions:

• sampleFromClient() helper exported for use in tools

• Specifies model preferences (prefers Claude Sonnet, falls back to any Claude)

• Gracefully no-ops if client doesn't support sampling

Architecture

src/
├── index.ts              # MCP server, capabilities, tool routing, elicitation, sampling
├── resources.ts          # Resources: list, read, subscribe, completion, pagination
├── tools/
│   ├── files.ts          # File system tools
│   ├── git.ts            # Git integration (simple-git)
│   ├── code-analysis.ts  # Static analysis tools
│   ├── http.ts           # HTTP client tools (axios)
│   └── system.ts         # OS/system tools
└── utils/
    ├── errors.ts         # Typed error classes, helpers
    └── logger.ts         # Structured log notifications (RFC 5424)

Key design decisions:

• Zod validation on every tool input — no silent failures

• Typed error classes with error codes for structured debugging

• Security: run_command has an explicit allowlist; env vars auto-mask secrets

• Cross-platform: works on Windows, macOS, and Linux

Tech Stack

• @modelcontextprotocol/sdk — MCP protocol

• simple-git — Git operations

• axios — HTTP client

• fast-glob — File pattern matching

• zod — Runtime schema validation

• TypeScript (strict mode)

License

MIT