HugeContext

Dynamic, Intent-Aware Semantic Memory for LLMs

HugeContext is a VS Code extension and standalone MCP server that acts as a local Context Engine for AI assistants. It provides intelligent, contextual code understanding for Claude Desktop, Cursor, Kilo Code, and other MCP-compatible tools.

Unlike cloud-based solutions, HugeContext runs entirely locally - your code never leaves your machine.

Why HugeContext (for everyday use)

• Local & private: Everything runs on your machine; nothing leaves your repo.
• Works with your tools: VS Code + MCP clients (Claude Desktop, Cursor, Gemini Code Assist, etc.).
• Instant answers: Understands code, docs, tests, and recent edits.
• Smart ranking: Intent-aware (debug, refactor, understand, tests) so you get the right snippets first.
• Zero setup after install: Opens, indexes, and keeps up with your edits automatically.

Quick Start (2 minutes, user-friendly)

Step 1: Install the Extension

code --install-extension hugecontext-1.3.2.vsix

Or install via VS Code: Extensions → ... → Install from VSIX...

Step 2: Open Your Project

Open any project folder in VS Code. HugeContext will:

1. Initialize the built-in CodeRankEmbed Q8 model (included in VSIX)
2. Index your codebase automatically (semantic graph + AST)
3. Watch for changes in real-time

Step 3: Open the Dashboard

Press Cmd/Ctrl+Shift+P → HugeContext: Show Dashboard

The Dashboard shows:

• Indexing status and statistics
• One-click MCP setup for popular AI clients
• Quick actions

MCP Integration (AI Assistants)

HugeContext exposes a query_context tool via the Model Context Protocol (MCP) so your AI assistant can pull in the right code snippets automatically.

Supported Clients

Global Configuration (Recommended)

For clients like Kilo Code that support global MCP settings, use this config:

{
  "mcpServers": {
    "hugecontext": {
      "command": "node",
      "args": ["/path/to/HugeContext/dist/mcp-cli.sec.js"]
    }
  }
}

Important: Don't include --workspace in global config. The MCP server automatically detects the workspace from:

1. The workspace parameter in each query (if provided)
2. The current working directory (cwd)

Cursor / Windsurf / Antigravity (VSIX + MCP)

1. Install the VSIX in your IDE: "Install from VSIX…" (or CLI e.g. cursor --install-extension hugecontext-1.3.2.vsix).

2. Add the MCP block to the IDE MCP settings. Use the global shim path (recommended):

   {
     "mcpServers": {
       "hugecontext": {
         "command": "/Users/YOUR_USER/.hugecontext/bin/hugecontext-mcp"
       }
     }
   }
   

   Note: Replace YOUR_USER with your system username.

3. Open a project; HugeContext automatically detects the workspace using:

   • Environment variables (if set by the IDE)
   • Workspace pointer file (~/.hugecontext/codex_workspace) - automatically updated by the extension
   • Current working directory (fallback)

   The extension updates the workspace pointer file whenever you switch projects, so the MCP server always knows which workspace to use.

4. Use the dashboard inside the IDE (HugeContext: Show Dashboard) to verify indexing and MCP.

Multi-Workspace Support

HugeContext v1.3.2+ supports dynamic workspace switching:

• One global MCP config works with ALL your projects
• Server automatically switches databases when you change repos
• Workspace pointer file: The VS Code extension automatically updates ~/.hugecontext/codex_workspace with your current workspace, enabling zero-config detection.

Workspace Detection Priority:

1. workspace parameter in query_context (explicit override)
2. CODEX_WORKSPACE_ROOT environment variable
3. WORKSPACE_PATH environment variable
4. Workspace pointer file (~/.hugecontext/codex_workspace) - updated by extension
5. Current working directory (cwd)

The AI assistant can explicitly specify a workspace:

{
  "query": "how does authentication work?",
  "workspace": "/Users/you/projects/my-app"
}

Or let it auto-detect using the workspace pointer file (recommended when using the VS Code extension).

How It Works

Intent-Aware Retrieval

HugeContext automatically detects what you're trying to do and adjusts its search strategy:

Quality-First Results

Unlike simple search, HugeContext prioritizes quality over quantity:

• Minimum similarity threshold (0.25) - Only returns genuinely relevant results
• Score cliff detection - Stops returning results when quality drops significantly
• Test file filtering - Deprioritizes tests for non-test queries
• Dynamic content length - More content for highly relevant files

Semantic Search

Uses CodeRankEmbed Q8 (Quantized ONNX):

• High-Performance: Q8 quantization for faster local inference.
• 768 Dimensions: High semantic resolution for complex code patterns.
• Embedded: Included in the VSIX for private, offline use.

Code Graph

Builds a hypergraph of relationships:

• File → Function/Class → Method hierarchy
• Import/export dependencies
• Test ↔ Implementation connections
• Git co-change patterns

Heat Tracking

Real-time developer activity tracking:

• Recently viewed/edited files get higher "heat"
• Heat propagates through graph connections
• Prioritizes contextually relevant code

VS Code Commands

Configuration

VS Code Settings

Supported Languages

Changelog

v1.0.17

• Multi-workspace support: Single global MCP config works with all projects
• Dynamic DB switching: Server automatically loads correct database per query
• Workspace parameter: Optional workspace param in query_context

v1.0.15-16

• MCP transport fix: Resolved JSON-RPC errors from log output
• Improved DB selection: Only loads database matching current workspace

v1.0.13-14

• Quality-first filtering: Minimum similarity threshold, score cliff detection
• Reduced noise: Tests deprioritized for non-test queries
• Git history optimization: Removed from non-debug pipelines
• Dynamic content length: More content for highly relevant results

v1.0.12

• Snowflake Arctic Embed: Upgraded to optimized code search model
• CLS pooling: Better embedding quality for code

License

HugeContext Internal Use License - see LICENSE for details.

Usage: Free for personal and internal business use. Restrictions: No redistribution, sublicensing, or making available to third parties.

Acknowledgments

• HugeContext is created and maintained by J. David Garcia (@codavidgarcia).
• Snowflake Arctic Embed - Code-optimized embeddings
• Tree-sitter - Incremental parsing
• Transformers.js - ONNX inference
• sql.js - Pure JS SQLite
• Model Context Protocol - AI integration standard