Thanks to visit codestin.com
Credit goes to github.com

Skip to content
/ smart-coding-mcp Public

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

License

MIT license
94 stars 15 forks Branches Tags Activity
Star
Notifications You must be signed in to change notification settings

omar-haris/smart-coding-mcp

BranchesTags

Repository files navigation

Smart Coding MCP

npm version npm downloads License: MIT Node.js

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

Tool Description Example
semantic_search Find code by meaning, not just keywords "Where do we validate user input?"
index_codebase Manually trigger reindexing Use after major refactoring or branch switches
clear_cache Reset the embeddings cache Useful when cache becomes corrupted
d_check_last_version Get latest version of any package (20 ecosystems) "express", "npm:react", "pip:requests"
e_set_workspace Change project path at runtime Switch to different project without restart
f_get_status Get server info: version, index status, config Check indexing progress, model info, cache size

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.

Example

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:

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:

IDE OS Config Path
Claude Desktop macOS ~/Library/Application Support/Claude/claude_desktop_config.json
Claude Desktop Windows %APPDATA%\Claude\claude_desktop_config.json
Cascade (Cursor) All Configured via UI Settings > Features > MCP
Antigravity macOS ~/.gemini/antigravity/mcp_config.json
Antigravity Windows %USERPROFILE%\.gemini\antigravity\mcp_config.json

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}"]
    }
  }
}
Client Supports ${workspaceFolder}
VS Code Yes
Cursor (Cascade) Yes
Antigravity No ❌
Claude Desktop No ❌

Environment Variables

Override configuration settings via environment variables in your MCP config:

Variable Type Default Description
SMART_CODING_VERBOSE boolean false Enable detailed logging
SMART_CODING_BATCH_SIZE number 100 Files to process in parallel
SMART_CODING_MAX_FILE_SIZE number 1048576 Max file size in bytes (1MB)
SMART_CODING_CHUNK_SIZE number 25 Lines of code per chunk
SMART_CODING_MAX_RESULTS number 5 Max search results
SMART_CODING_SMART_INDEXING boolean true Enable smart project detection
SMART_CODING_WATCH_FILES boolean false Enable file watching for auto-reindex
SMART_CODING_SEMANTIC_WEIGHT number 0.7 Weight for semantic similarity (0-1)
SMART_CODING_EXACT_MATCH_BOOST number 1.5 Boost for exact text matches
SMART_CODING_EMBEDDING_MODEL string nomic-ai/nomic-embed-text-v1.5 AI embedding model to use
SMART_CODING_EMBEDDING_DIMENSION number 256 MRL dimension (64, 128, 256, 512, 768)
SMART_CODING_DEVICE string cpu Inference device (cpu, webgpu, auto)
SMART_CODING_CHUNKING_MODE string smart Code chunking (smart, ast, line)
SMART_CODING_WORKER_THREADS string auto Worker threads (auto or 1-32)

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

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["Vector Store<br/>embeddings.json"]
        Hashes["File Hashes<br/>Incremental updates"]
    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
Loading

Tech Stack

Component Technology
Protocol Model Context Protocol (JSON-RPC)
AI Model nomic-embed-text-v1.5 (MRL)
Inference transformers.js + ONNX Runtime
Chunking Smart regex / Tree-sitter AST
Search Cosine similarity + exact match boost

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.

About

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

Topics

javascript ai mcp ast gemini cursor codex rag mrl llm cluade local-rag antigravity

Resources

Readme

License

MIT license

Contributing

Contributing
Activity

Stars

94 stars

Watchers

2 watching

Forks

15 forks
Report repository

Releases

13 tags

Packages

No packages published

Languages