Multi-Agent MCP Orchestrator

An enterprise-grade, autonomous AI orchestration system built with LangGraph, utilizing the Model Context Protocol (MCP) to seamlessly connect LLMs to local environments, vector databases, and cloud APIs.

Architectural Highlights

1. The Supervisor Routing (LangGraph)

• Transitioned from a monolithic tool-calling agent to a distributed MultiServerMCPClient architecture.

• Implemented strict system prompt injection to control agent behavior dynamically (e.g., forcing Google Sheets defaults) without polluting user queries.

• Handled state routing via GraphState to allow seamless multi-step tool chaining across completely isolated MCP servers.

2. Self-Healing RAG Vault (ragsystem.py)

• Built a custom Pinecone Hybrid Search vector database with LlamaParse for structural markdown extraction.

• Wipe-and-Replace Mechanism: Implemented pre-ingestion metadata filtering (index.delete(filter={"source": filepath})). This prevents vector duplication when re-ingesting updated source files, maintaining a pristine context window.

3. Custom Google Workspace Server (workspace_custom.py)

• Bypassed limited community MCP packages to build a raw Python MCP server interacting directly with Google Docs and Sheets REST APIs.

• Capabilities: * create_sheet & write_sheet: Matrix-based (2D array) row appending using Google's USER_ENTERED parsing.

  • create_doc & append_doc: Dynamic EOF index calculation to securely inject text into heavily nested Google Docs JSON trees.

4. Cross-Platform Integrations

This orchestrator successfully routes complex workflows across multiple isolated domains in a single conversational turn:

• Local Filesystem: Read/Write access via standard I/O transport.

• Notion: Document retrieval via official @modelcontextprotocol/server-notion.

• DuckDuckGo: Zero-auth web search via duckduckgo-mcp-server.

• Todoist: Task management and creation via a custom FastMCP REST API server.

Related MCP server: Google Workspace MCP Server

Prerequisites & Installation

1. Environment Setup

This project uses uv for lightning-fast package management. Ensure Python 3.12+ and uv are installed.

# Clone the repository
git clone [https://github.com/yourusername/multi-agent-orchestrator.git](https://github.com/yourusername/multi-agent-orchestrator.git)
cd multi-agent-orchestrator

# Install dependencies via uv
uv venv
uv pip install -r requirements.txt