Smart Coding MCP

What is Smart Coding MCP about?

Smart Coding MCP enables AI coding assistants to retrieve relevant code fragments based on meaning rather than keyword matches. It builds a vector index of your project's source files using a local embedding model (nomic‑embed‑text‑v1.5) with flexible Matryoshka Representation Learning dimensions, then answers natural‑language queries with ranked code snippets.

How to use Smart Coding MCP?

1. Install the package globally (or run via npx).

   npm install -g smart-coding-mcp
   # or, one‑off execution
   npx -y smart-coding-mcp --workspace /path/to/project
   

2. Configure the MCP client (Claude Desktop, Antigravity, Cursor, etc.) to point to the server executable and workspace path as shown in the README.
3. Start the server; it begins indexing in the background while remaining responsive.
4. Invoke MCP tools such as semantic_search, index_codebase, clear_cache, e_set_workspace, or f_get_status from your IDE or assistant.

Key features of Smart Coding MCP

• Semantic Search: Retrieve code by concept, handling synonyms, typos, and different terminology.
• Flexible Embedding Dimensions: Choose 64‑768d embeddings; default 128d for speed/quality balance.
• Progressive & Incremental Indexing: Search works even while indexing; only changed files are re‑processed.
• Resource Throttling: CPU limited to 50 % by default, configurable batch size and delays.
• SQLite Cache: Fast binary storage with WAL for concurrency; automatic migration from JSON.
• Runtime Workspace Switching: Change project path without restarting the server.
• Comprehensive Status Reporting: Version, index progress, cache size, model info.
• Privacy‑First: All inference runs locally; no network calls or telemetry.

Use cases of Smart Coding MCP

• AI‑augmented pair programming: Assistants instantly locate validation logic, error handling, or specific API usages.
• Large monorepos: Quickly navigate across frontend/backend modules without scanning the whole tree each time.
• On‑premise development environments: Organizations that forbid external API calls can still benefit from semantic search.
• Educational tooling: Students ask natural‑language questions about sample projects and receive concrete code excerpts.

FAQ from Smart Coding MCP

Q: Do I need an internet connection? A: No. The embedding model is bundled and runs locally via ONNX Runtime.

Q: Can I use a different embedding model? A: Yes, set SMART_CODING_EMBEDDING_MODEL to another compatible model name.

Q: How does the server affect my laptop’s performance? A: By default it caps CPU usage at 50 % and processes files in configurable batches, keeping the system responsive.

Q: What languages are supported? A: The default configuration targets JavaScript/TypeScript, but the chunker can be extended to other languages.

Q: How do I reset the index? A: Use the clear_cache tool or delete the .smart-coding-cache/ folder and run index_codebase again.

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), with runtime workspace switching and comprehensive status reporting.

Available Tools

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

Performance & Resource Management

Progressive Indexing

• Search works immediately, even while indexing continues (like video buffering)
• Incremental saves every 5 batches - no data loss if interrupted
• Real-time indexing status shown when searching during indexing

Resource Throttling

• CPU usage limited to 50% by default (configurable)
• Your laptop stays responsive during indexing
• Configurable delays between batches
• Worker thread limits respect system resources

SQLite Cache

• 5-10x faster than JSON for large codebases
• Write-Ahead Logging (WAL) for better concurrency
• Binary blob storage for smaller cache size
• Automatic migration from JSON

Optimized Defaults

• 128d embeddings by default (2x faster than 256d, minimal quality loss)
• Smart batch sizing based on project size
• Parallel processing with auto-tuned worker threads

Installation

Install globally via npm:

npm install -g smart-coding-mcp

To update to the latest version:

npm update -g smart-coding-mcp

Configuration

Add to your MCP configuration file. The location depends on your IDE and OS:

Add the server configuration to the mcpServers object in your config file:

Option 1: Absolute Path (Recommended)

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

Option 2: Multi-Project Support

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

Option 3: Auto-Detection (May Not Work)

⚠️ Warning: Most MCP clients (including Antigravity and Claude Desktop) do NOT support ${workspaceFolder} variable expansion. The server will exit with an error if the variable is not expanded.

For clients that support dynamic variables (VS Code, Cursor):

{
  "mcpServers": {
    "smart-coding-mcp": {
      "command": "smart-coding-mcp",
      "args": ["--workspace", "${workspaceFolder}"]
    }
  }
}

Environment Variables

Override configuration settings via environment variables in your MCP config:

Example with environment variables:

{
  "mcpServers": {
    "smart-coding-mcp": {
      "command": "smart-coding-mcp",
      "args": ["--workspace", "/path/to/project"],
      "env": {
        "SMART_CODING_VERBOSE": "true",
        "SMART_CODING_BATCH_SIZE": "200",
        "SMART_CODING_MAX_FILE_SIZE": "2097152"
      }
    }
  }
}

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

How It Works

Tech Stack

Search Flow

Query → Vector embedding → Cosine similarity → Ranked results

Examples

Natural language search:

Query: "How do we handle cache persistence?"

Result:

// lib/cache.js (Relevance: 38.2%)
async save() {
  await fs.writeFile(cacheFile, JSON.stringify(this.vectorStore));
  await fs.writeFile(hashFile, JSON.stringify(this.fileHashes));
}

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.

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: nomic-embed-text-v1.5 via transformers.js v3

• Matryoshka Representation Learning (MRL) for flexible dimensions
• Configurable output: 64, 128, 256, 512, or 768 dimensions
• Longer context (8192 tokens vs 256 for MiniLM)
• Better code understanding through specialized training
• WebGPU support for up to 100x faster inference (when available)

Legacy Model: all-MiniLM-L6-v2 (fallback)

• Fast inference, small footprint (~100MB)
• Fixed 384-dimensional output

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

License

MIT License

Copyright (c) 2025 Omar Haris

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.