Local Search MCP Server

A Model Context Protocol (MCP) server that enables AI assistants to perform semantic search across indexed documents using vector embeddings. Index documents from GitHub repositories and URLs to power natural language queries with contextual results.

Table of Contents

• Features

• Installation

  • Prerequisites

  • Setup

• Configuration

• Usage

• Tools

• Documentation

• Development

• Contributing

• Author

• License

Features

• Semantic Search: Natural language queries over indexed documents using transformer embeddings

• Repository Indexing: Clone and index GitHub repositories with configurable file patterns

• File Downloads: Fetch and index files from URLs with automatic processing

• Async Processing: Non-blocking operations with job progress tracking

• SQLite Storage: Efficient vector storage with optimized similarity search

• MCP Protocol: Full compatibility with Claude Desktop and other MCP applications

Quick Start

The fastest way to get started is using npx (no cloning or building required):

MCP Configuration (npx)

Add to Claude Desktop's claude_desktop_config.json:

Installation

Prerequisites

• Node.js >= 18.0.0

• npm or yarn package manager

• Git for cloning repositories (development only)

Option 1: NPM Package (Recommended)

Option 2: From Source (Development)

MCP Configuration

For NPM package installation:

For source installation:

Usage

Once configured, the server provides semantic search capabilities within Claude Desktop and other MCP-compatible applications.

Tools

The Local Search MCP Server provides 7 tools for document indexing and semantic search:

🔍 Search Tools

search_documents

Perform AI-enhanced semantic search with content classification, domain detection, and intelligent recommendations.

Parameters:

• query (required): Natural language search query

• options (optional): Search configuration object

  • limit (number, default: 10): Maximum results to return

  • minScore (number, default: 0.7): Minimum similarity score (0-1)

  • includeMetadata (boolean, default: true): Include metadata in results

  • domainFilter (array): Filter by technology domains (e.g., ["javascript", "python"])

  • contentTypeFilter (array): Filter by content type ("code", "docs", "config", "mixed")

  • languageFilter (array): Filter by programming language (e.g., ["typescript", "javascript"])

  • minQualityScore (number): Minimum content quality score (0-1)

  • minAuthorityScore (number): Minimum source authority score (0-1)

Example:

get_file_details

Retrieve detailed content of a specific file with surrounding chunk context.

Parameters:

• filePath (required): Absolute path to file

• chunkIndex (optional): Specific chunk to retrieve with surrounding context

• contextSize (number, default: 3): Number of chunks to include before and after target chunk

📦 Content Management Tools

fetch_repo

Clone a Git repository (GitHub, Azure DevOps, etc.) using repomix, convert to markdown, and add to searchable index. Returns job ID for progress tracking.

Parameters:

• repoUrl (required): Git repository URL

• branch (optional): Branch/tag/commit, defaults to main/master

• options (optional): Repository processing options

  • includePatterns (array, default: ["/*.md", "/.mdx", "**/.txt", "/*.json", "/.rst", "**/.yml", "**/*.yaml"]): File patterns to include

  • excludePatterns (array, default: ["/node_modules/"]): File patterns to exclude

  • outputStyle (string, default: "markdown"): Output format (fixed to markdown)

  • removeComments (boolean, default: false): Remove comments from code files

  • showLineNumbers (boolean, default: true): Show line numbers in output

Example:

fetch_file

Download a single file from a URL and add it to the searchable index. Returns job ID for progress tracking.

Parameters:

• url (required): URL of file to download

• filename (required): Desired filename for saving

• options (optional): Download options

  • overwrite (boolean, default: true): Whether to overwrite existing files

  • indexAfterSave (boolean, default: true): Automatically index after download

  • maxFileSizeMB (number, default: 1024): Maximum file size in MB

remove_file

Delete a file and all its associated chunks and embeddings from the index.

Parameters:

• filePath (required): Absolute path to file to remove

flush_all

Flush the entire database and all downloaded files. WARNING: This action is irreversible and will delete all indexed content, documents, and cached files.

Parameters: None

What gets deleted:

• All vector embeddings and document chunks from the database

• All recommendation and learning data

• All downloaded files from the fetched directory

• All cloned repositories from the repositories directory

• All temporary files from the temp directory

• All active background jobs are cancelled

Example:

⚙️ Job Management Tools

get_job_status

Get status and progress of an async job by ID with real-time accurate progress.

Parameters:

• jobId (required): Job ID returned from fetch_* operations

Returns:

• Job status: "running", "completed", or "failed"

• Progress percentage (0-100)

• Duration and timestamps

• Error message if failed

• Result data if completed

list_active_jobs

List all currently active (running) jobs with their status and progress.

Parameters: None

Returns:

• List of active jobs with progress

• Job statistics (total, completed, failed, average duration)

• Real-time progress updates

Documentation

For detailed technical documentation:

• Architecture - System design and processing pipeline

• API Reference - Complete tool specifications and types

• Performance - Optimization guides and benchmarks

• Usage Examples - Sample integrations and workflows

Development

Configuration

Environment Variables

Set optional environment variables for custom paths:

• MCP_DATA_FOLDER - Custom database and logs directory (defaults to platform-specific user data folder)

• MCP_DOCS_FOLDER - Custom document storage directory (defaults to platform-specific user documents folder)

Supported File Types

The server processes these file types:

• Documentation: .md, .txt, .rst, .yaml, .yml

• Data: .json, .csv

• Code: .js, .ts, .py, .java, .c, .cpp, .html, .css

• Files up to 1GB are supported

Acknowledgments

• @xenova/transformers - JavaScript ML models for embeddings

• sqlite-vec - Native vector search in SQLite

• better-sqlite3 - Fast SQLite3 bindings

• Model Context Protocol - Standard for AI tool integration

• repomix - Repository processing utility

Release Process

This project uses automated semantic versioning and publishing through GitHub Actions and semantic-release.

Commit Message Format

Follow Conventional Commits specification:

Types that trigger releases:

• feat: - New features (minor version bump)

• fix: - Bug fixes (patch version bump)

• perf: - Performance improvements (patch version bump)

• BREAKING CHANGE: - Breaking changes (major version bump)

Other types (no release):

• docs: - Documentation changes

• style: - Code formatting

• refactor: - Code refactoring

• test: - Adding tests

• chore: - Build process or auxiliary tool changes

Contributing

1. Fork the repository

2. Create a feature branch with descriptive name

3. Make changes following conventional commit format

4. Submit a pull request targeting the main branch

5. Ensure all CI checks pass before requesting review

Author

Patrick Ruddiman
GitHub

License

MIT - see LICENSE for details.