Smart Coding MCP

An extensible Model Context Protocol (MCP) server that provides intelligent semantic code search for AI assistants. Built with local AI models using Matryoshka Representation Learning (MRL) for flexible embedding dimensions (64-768d).

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.

Available Tools

🔍 a_semantic_search - Find Code by Meaning

The primary tool for codebase exploration. Uses AI embeddings to understand what you're looking for, not just match keywords.

How it works: Converts your natural language query into a vector, then finds code chunks with similar meaning using cosine similarity + exact match boosting.

Best for:

• Exploring unfamiliar codebases: "How does authentication work?"
• Finding related code: "Where do we validate user input?"
• Conceptual searches: "error handling patterns"
• Works even with typos: "embeding modle initializashun" still finds embedding code

Example queries:

"Where do we handle cache persistence?"
"How is the database connection managed?"
"Find all API endpoint definitions"

📦 d_check_last_version - Package Version Lookup

Fetches the latest version of any package from its official registry. Supports 20+ ecosystems.

How it works: Queries official package registries (npm, PyPI, Crates.io, etc.) in real-time. No guessing, no stale training data.

Supported ecosystems: npm, PyPI, Crates.io, Maven, Go, RubyGems, NuGet, Packagist, Hex, pub.dev, Homebrew, Conda, and more.

Best for:

• Before adding dependencies: "express" → 4.18.2
• Checking for updates: "pip:requests" → 2.31.0
• Multi-ecosystem projects: "npm:react", "go:github.com/gin-gonic/gin"

Example usage:

"What's the latest version of lodash?"
"Check if there's a newer version of axios"

🔄 b_index_codebase - Manual Reindexing

Triggers a full reindex of your codebase. Normally not needed since indexing is automatic and incremental.

How it works: Scans all files, generates new embeddings, and updates the SQLite cache. Uses progressive indexing so you can search while it runs.

When to use:

• After major refactoring or branch switches
• After pulling large changes from remote
• If search results seem stale or incomplete
• After changing embedding configuration (dimension, model)

🗑️ c_clear_cache - Reset Everything

Deletes the embeddings cache entirely, forcing a complete reindex on next search.

How it works: Removes the .smart-coding-cache/ directory. Next search or index operation starts fresh.

When to use:

• Cache corruption (rare, but possible)
• Switching embedding models or dimensions
• Starting fresh after major codebase restructure
• Troubleshooting search issues

📂 e_set_workspace - Switch Projects

Changes the workspace path at runtime without restarting the server.

How it works: Updates the internal workspace reference, creates cache folder for new path, and optionally triggers reindexing.

When to use:

• Working on multiple projects in one session
• Monorepo navigation between packages
• Switching between related repositories

ℹ️ f_get_status - Server Health Check

Returns comprehensive status information about the MCP server.

What it shows:

• Server version and uptime
• Workspace path and cache location
• Indexing status (ready, indexing, percentage complete)
• Files indexed and chunk count
• Model configuration (name, dimension, device)
• Cache size and type

When to use:

• Start of session to verify everything is working
• Debugging connection or indexing issues
• Checking indexing progress on large codebases

Installation

npm install -g smart-coding-mcp

To update:

npm update -g smart-coding-mcp

IDE Integration

Detailed setup instructions for your preferred environment:

Quick Setup

Add to your MCP config file:

{
  "mcpServers": {
    "smart-coding-mcp": {
      "command": "smart-coding-mcp",
      "args": ["--workspace", "/absolute/path/to/your/project"]
    }
  }
}

Config File Locations

Multi-Project Setup

{
  "mcpServers": {
    "smart-coding-frontend": {
      "command": "smart-coding-mcp",
      "args": ["--workspace", "/path/to/frontend"]
    },
    "smart-coding-backend": {
      "command": "smart-coding-mcp",
      "args": ["--workspace", "/path/to/backend"]
    }
  }
}

Environment Variables

Customize behavior via environment variables:

Example with env vars:

{
  "mcpServers": {
    "smart-coding-mcp": {
      "command": "smart-coding-mcp",
      "args": ["--workspace", "/path/to/project"],
      "env": {
        "SMART_CODING_VERBOSE": "true",
        "SMART_CODING_MAX_RESULTS": "10",
        "SMART_CODING_EMBEDDING_DIMENSION": "256"
      }
    }
  }
}

Performance

Progressive Indexing - Search works immediately while indexing continues in the background. No waiting for large codebases.

Resource Throttling - CPU limited to 50% by default. Your machine stays responsive during indexing.

SQLite Cache - 5-10x faster than JSON. Automatic migration from older JSON caches.

Incremental Updates - Only changed files are re-indexed. Saves every 5 batches, so no data loss if interrupted.

Optimized Defaults - 128d embeddings (2x faster than 256d with minimal quality loss), smart batch sizing, parallel processing.

How It Works

flowchart TB
    subgraph IDE["IDE / AI Assistant"]
        Agent["AI Agent<br/>(Claude, GPT, Gemini)"]
    end

    subgraph MCP["Smart Coding MCP Server"]
        direction TB
        Protocol["Model Context Protocol<br/>JSON-RPC over stdio"]
        Tools["MCP Tools<br/>semantic_search | index_codebase | set_workspace | get_status"]

        subgraph Indexing["Indexing Pipeline"]
            Discovery["File Discovery<br/>glob patterns + smart ignore"]
            Chunking["Code Chunking<br/>Smart (regex) / AST (Tree-sitter)"]
            Embedding["AI Embedding<br/>transformers.js + ONNX Runtime"]
        end

        subgraph AI["AI Model"]
            Model["nomic-embed-text-v1.5<br/>Matryoshka Representation Learning"]
            Dimensions["Flexible Dimensions<br/>64 | 128 | 256 | 512 | 768"]
            Normalize["Layer Norm → Slice → L2 Normalize"]
        end

        subgraph Search["Search"]
            QueryEmbed["Query → Vector"]
            Cosine["Cosine Similarity"]
            Hybrid["Hybrid Search<br/>Semantic + Exact Match Boost"]
        end
    end

    subgraph Storage["Cache"]
        Vectors["SQLite Database<br/>embeddings.db (WAL mode)"]
        Hashes["File Hashes<br/>Incremental updates"]
        Progressive["Progressive Indexing<br/>Search works during indexing"]
    end

    Agent <-->|"MCP Protocol"| Protocol
    Protocol --> Tools

    Tools --> Discovery
    Discovery --> Chunking
    Chunking --> Embedding
    Embedding --> Model
    Model --> Dimensions
    Dimensions --> Normalize
    Normalize --> Vectors

    Tools --> QueryEmbed
    QueryEmbed --> Model
    Cosine --> Hybrid
    Vectors --> Cosine
    Hybrid --> Agent

Tech Stack

Privacy

Everything runs 100% locally:

• AI model runs on your machine (no API calls)
• Code never leaves your system
• No telemetry or analytics
• Cache stored in .smart-coding-cache/

Research Background

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

License

MIT License - Copyright (c) 2025 Omar Haris

See LICENSE for full text.