Smart Coding MCP

An extensible Model Context Protocol (MCP) server that provides intelligent semantic code search for AI assistants. Built with local AI models, inspired by Cursor's semantic search research.

What This Does

AI coding assistants work better when they can find relevant code quickly. Traditional keyword search falls short - if you ask "where do we handle authentication?" but your code uses "login" and "session", keyword search misses it.

This MCP server solves that by indexing your codebase with AI embeddings. Your AI assistant can search by meaning instead of exact keywords, finding relevant code even when the terminology differs.

Why Use This

Better Code Understanding

• Search finds code by concept, not just matching words

• Works with typos and variations in terminology

• Natural language queries like "where do we validate user input?"

Performance

• Pre-indexed embeddings are faster than scanning files at runtime

• Smart project detection skips dependencies automatically (node_modules, vendor, etc.)

• Incremental updates - only re-processes changed files

Privacy

• Everything runs locally on your machine

• Your code never leaves your system

• No API calls to external services

Installation

Install globally via npm:

Configuration

Add to your MCP configuration file (e.g., ~/.config/claude/mcp.json or similar):

Option 1: Specific Project (Recommended)

Option 2: Multi-Project Support

Option 3: Auto-Detect Current Directory

Environment Variables

Override configuration settings via environment variables in your MCP config:

Example with environment variables:

Note: The server starts instantly and indexes in the background, so your IDE won't be blocked waiting for indexing to complete.

Available Tools

semantic_search - Find code by meaning

index_codebase - Manually trigger reindexing

clear_cache - Reset the embeddings cache

How It Works

The server indexes your code in four steps:

1. Discovery: Scans your project for source files

2. Chunking: Breaks code into meaningful pieces (respecting function boundaries)

3. Embedding: Converts each chunk to a vector using a local AI model

4. Storage: Saves embeddings to .smart-coding-cache/ for fast startup

When you search, your query is converted to the same vector format and compared against all code chunks using cosine similarity. The most relevant matches are returned.

Smart Project Detection

The server detects your project type by looking for marker files and automatically applies appropriate ignore patterns:

JavaScript/Node (package.json found)

• Ignores: node_modules, dist, build, .next, coverage

Python (requirements.txt or pyproject.toml)

• Ignores: pycache, venv, .pytest_cache, .tox

Android (build.gradle)

• Ignores: .gradle, build artifacts, generated code

iOS (Podfile)

• Ignores: Pods, DerivedData, xcuserdata

And more: Go, PHP, Rust, Ruby, .NET

This typically reduces indexed file count by 100x. A project with 50,000 files (including node_modules) indexes just 500 actual source files.

Configuration

The server works out of the box with sensible defaults. Create a config.json file in your workspace to customize:

Key options:

• smartIndexing: Enable automatic project type detection and smart ignore patterns (default: true)

• verbose: Show detailed indexing logs (default: false)

• watchFiles: Automatically reindex when files change (default: true)

• enableCache: Cache embeddings to disk (default: true)

• chunkSize: Lines of code per chunk - smaller = more precise, larger = more context (default: 15)

• batchSize: Number of files to process in parallel (default: 100)

• maxFileSize: Skip files larger than this size in bytes (default: 1MB)

Examples

Natural language search:

Query: "How do we handle cache persistence?"

Result:

Typo tolerance:

Query: "embeding modle initializashun"

Still finds embedding model initialization code despite multiple typos.

Conceptual search:

Query: "error handling and exceptions"

Finds all try/catch blocks and error handling patterns.

Performance

Tested on a typical JavaScript project:

Supported File Types

Languages: JavaScript, TypeScript, Python, Java, Kotlin, Scala, C, C++, C#, Go, Rust, Ruby, PHP, Swift, Shell

Web: HTML, CSS, SCSS, Sass, XML, SVG

Config/Data: JSON, YAML, TOML, SQL

Total: 36 file extensions

Architecture

The modular design makes it easy to add new features. See ARCHITECTURE.md for implementation details.

Troubleshooting

"Server can't find config.json"

Make sure cwd is set in your MCP configuration to the full path of smart-coding-mcp.

"Indexing takes too long"

• Verify smartIndexing is enabled

• Add more patterns to excludePatterns

• Reduce fileExtensions to only what you need

"Search results aren't relevant"

• Try more specific queries

• Increase maxResults to see more options

• Run index_codebase to force a full reindex

"Cache corruption errors"

Use the clear_cache tool or run:

CLI Commands

Privacy

• AI model runs entirely on your machine

• No network requests to external services

• No telemetry or analytics

• Cache stored locally in .smart-coding-cache/

Technical Details

Embedding Model: all-MiniLM-L6-v2 via transformers.js

• Fast inference (CPU-friendly)

• Small model size (~100MB)

• Good accuracy for code search

Vector Similarity: Cosine similarity

• Efficient comparison of embeddings

• Normalized vectors for consistent scoring

Hybrid Scoring: Combines semantic similarity with exact text matching

• Semantic weight: 0.7 (configurable)

• Exact match boost: 1.5x (configurable)

Research Background

This project builds on research from Cursor showing that semantic search improves AI coding agent performance by 12.5% on average across question-answering tasks. The key insight is that AI assistants benefit more from relevant context than from large amounts of context.

See: https://cursor.com/blog/semsearch

Contributing

Contributions are welcome. See CONTRIBUTING.md for guidelines.

Potential areas for improvement:

• Additional language support

• Code complexity analysis

• Refactoring pattern detection

• Documentation generation

License

MIT - see LICENSE file

Documentation

• ARCHITECTURE.md - Implementation details and design decisions

• CONTRIBUTING.md - Guidelines for contributors

• EXAMPLES.md - More usage examples

• QUICKSTART.md - Detailed setup guide