---

Omniscience is a highly optimized Model Context Protocol (MCP) server designed to give Large Language Models (LLMs) token-efficient, surgical access to massive codebases. Instead of flooding the LLM's context window with entire repositories, Omniscience uses a sophisticated Dual-Brain architecture to find exactly what the LLM needs—and absolutely nothing more.

🧠 The Dual-Brain Architecture

graph TD
    A[Codebase] -->|Real-time watcher| B(Omniscience Scanner)
    B -->|Code| C{Dual-Brain Parser}
    
    subgraph Structural Brain
    C -->|AST Parsing| D[Tree-Sitter]
    D -->|Function Definitions & Calls| E[(SQLite Graph DB)]
    end
    
    subgraph Semantic Brain
    C -->|Text/Code| F[Voyage-4-nano]
    F -->|Local Embeddings| G[(LanceDB Vector DB)]
    end
    
    E -.->|Graph Query| H[MCP Client]
    G -.->|Semantic Search| H

1. Structural Brain (Tree-sitter)

Parses the AST (Abstract Syntax Tree) of your codebase in real-time. It maps out exact file locations, boundary lines for functions/classes, and automatically generates a complete Call-Graph (Caller -> Callee relationships) stored in a local SQLite database.

2. Semantic Brain (LanceDB & Voyage-4-nano)

Generates and stores high-quality semantic embeddings of every code symbol completely locally. Allows the LLM to search for abstract concepts ("how does the auth routing work?") using lightning-fast hybrid search.

📖 How to talk to your AI?

If you're wondering how exactly you should prompt your AI (Claude, Antigravity, Cursor) to make use of these superpowers, check out our Prompt Library (PROMPTS.md) for copy-pasteable examples!

---

🛠️ Exposed MCP Tools

The server exposes powerful tools to the AI, allowing it to navigate your project like a senior engineer.

Tool	Description	Token Impact
🔍 semantic_search	Finds relevant code symbols based on a natural language query or keywords.	Low
🕸️ graph_query	Returns the blast radius of a specific symbol based on the AST Call-Graph.	Low
📖 surgical_read	Extracts only the exact code snippet for a single function or class.	Massive Savings
🏗️ apply_surgical_patch	Replaces an exact code symbol with new code and triggers a background re-index.	Low
🔄 rebuild_index	Manually triggers a complete re-indexing of the entire workspace.	None

---

🚀 Installation & Setup

Omniscience is designed to be ridiculously fast. We use uv for lightning-fast dependency resolution.

# 1. Clone the repository
git clone https://github.com/FreakyLetsFail/mcp-omniscience.git
cd mcp-omniscience

# 2. Run the Initialization Script (Downloads model, syncs env)
./init.sh

📦 Standalone CLI Indexer (For Large Repositories)

To prevent your IDE and OS from freezing when opening a massive repository for the first time, Omniscience comes with a standalone CLI tool. It builds the AST Call-Graph and Semantic Vector Database efficiently in the background before you even start your AI.

./index.sh index /path/to/your/large/project

This creates a .omniscience folder directly inside your project containing the LanceDB and SQLite databases.

🔌 IDE Integration

Add Omniscience to your MCP client configuration (mcp_config.json, claude_desktop_config.json, etc.):

{
  "mcpServers": {
    "omniscience": {
      "command": "/path/to/mcp-omniscience/run_server.sh",
      "args": []
    }
  }
}

[!TIP] No initialization prompt required! When the MCP server starts in a new WORKSPACE_DIR, it automatically builds the vector and graph databases in the background.

---

💰 Token Cost Analysis

Why use Omniscience over traditional whole-file reading?

• Full File Read (server.py): ~911 Tokens
• Omniscience Surgical Read (1 function): ~117 Tokens
• Context Window Saved: 87.16% per interaction!

By isolating exactly what is needed, the LLM hallucinates less, replies faster, and drastically reduces API costs.

---