Skip to content

MCP_SERVER.md

Latest commit

 

History

History
198 lines (148 loc) · 4.19 KB

MCP_SERVER.md

File metadata and controls

198 lines (148 loc) · 4.19 KB

🚀 Blink MCP Server

MCP (Model Context Protocol) Server for the Blink AI Agent - Enables Claude to autonomously access files, explore projects, and understand code structure.

What is MCP?

MCP is a protocol that lets AI models (Claude) call tools/functions to interact with external systems. Instead of just sending text, Claude can:

  • ✅ Read files when it needs context
  • ✅ Explore project structure
  • ✅ Search for related code
  • ✅ Understand file metadata

Architecture

User Input
    ↓
CLI (simplified_cli.py)
    ↓
EnhancedCodeAgent (with MCP)
    ↓
BlinkMCPServer (provides tools)
    ↓
Claude API (uses tools autonomously)
    ↓
Generated Code (TypeScript/Python/etc)

Available MCP Tools

1. read_file

Read the full contents of a file

Tool: read_file
Input: {"path": "src/services/sensor.ts"}
Output: {
  "success": true,
  "language": "TypeScript",
  "size": 1024,
  "lines": 45,
  "content": "..."
}

2. list_directory

List files and folders in a path

Tool: list_directory
Input: {"path": "src/modules"}
Output: {
  "success": true,
  "directories": ["sensors", "controllers"],
  "files": ["index.ts", "config.ts"],
  "file_count": 2,
  "dir_count": 2
}

3. search_files

Search for files matching a pattern

Tool: search_files
Input: {"pattern": "sensor", "directory": "src"}
Output: {
  "success": true,
  "matches": ["src/services/temperature-sensor.ts", "src/services/ph-sensor.ts"],
  "count": 2
}

4. get_file_info

Get metadata without reading full content

Tool: get_file_info
Input: {"path": "src/services/sensor.ts"}
Output: {
  "success": true,
  "language": "TypeScript",
  "size": 1024,
  "size_kb": 1.0,
  "lines": 45,
  "extension": ".ts"
}

5. get_project_structure

Get overview of project organization

Tool: get_project_structure
Input: {"max_depth": 3}
Output: {
  "success": true,
  "structure": {
    "src/": {
      "services/": {...},
      "models/": {...}
    }
  }
}

How It Works

Before MCP (Old Way):

  1. You tell the CLI: "Create pH sensor from temperature sensor"
  2. CLI reads temperature-sensor.ts, embeds full content in prompt
  3. Claude receives: "Here's the full file... generate code..."
  4. Problem: Claude doesn't know it's TypeScript, generates Python

With MCP (New Way):

  1. You tell the CLI: "Create pH sensor from temperature sensor"
  2. CLI tells Claude about MCP tools available
  3. Claude intelligently calls read_file to examine the temperature-sensor.ts
  4. Claude sees it's TypeScript, understands the patterns
  5. Claude generates TypeScript code that matches the original

Benefits

✅ Context-aware - Claude reads what it needs, when it needs it ✅ Language-aware - Claude understands file types and generates correct language ✅ Token-efficient - Only reads relevant parts, saves API costs ✅ Scalable - Works with large codebases ✅ Smart - Claude can explore dependencies and related files

Usage

Using the MCP Server Directly:

from src.enhanced_agent import EnhancedCodeAgent

agent = EnhancedCodeAgent()

# Get tools available
tools = agent.get_mcp_tools()

# Call a tool directly
result = agent.call_mcp_tool("read_file", {"path": "src/app.ts"})

Through the CLI:

python main.py

Then use generate:: command as usual:

generate:: create pH sensor version from "src/temperature-sensor.ts" like "src/ph-sensor.ts"

The CLI will:

  1. Use MCP to get file info
  2. Pass enhanced instruction to Claude with MCP tools
  3. Claude uses tools to examine files
  4. Generate correct TypeScript code

Implementation Files

  • src/mcp_server.py - BlinkMCPServer class with all tools
  • src/enhanced_agent.py - EnhancedCodeAgent that uses MCP
  • src/simplified_cli.py - Updated CLI with MCP integration

Testing

# Test MCP server
python test_mcp.py

# Use enhanced CLI
python main.py

Next Steps

The MCP server is fully functional and integrated. Claude now:

  • ✅ Has tools to autonomously access files
  • ✅ Understands project structure
  • ✅ Generates code in the correct language
  • ✅ Uses context intelligently

Just use python main.py as normal - the MCP magic happens behind the scenes!