mcp-agent-graph

MCP Agent Graph is a multi-agent system that combines agents, long-term memory, and an MCP-friendly workflow engine into a visual development experience. It enables you to design workflows with a drag-and-drop Graph Editor, connect external tools and data sources via the MCP protocol, and manage agents that can reason, fetch tools, and store context across sessions. The platform exposes components for graph orchestration, memory, prompts, and collaboration, making it easier to build, test, and deploy complex agent-based applications. You can explore the core features such as the Visual Graph Editor, Subgraph nesting, smart handoffs, and task scheduling, all of which work together through the MCP protocol to integrate external services and tools.

To use the server, follow the deployment guide to start the backend, then access the frontend and API endpoints. The backend provides endpoints for health checks and API documentation, while the frontend (if you enable it) offers a UI to design graphs, manage agents, and decorate prompts. Keep in mind the project status: the repository notes that the project is paused and an online demo service has been discontinued, so you should rely on local deployment for testing and development.

Prerequisites:

• Git installed
• Python 3.11+ installed (or a Python environment with venv)
• Docker and Docker Compose (optional for Docker deployment)
• Optional: Node.js if you intend to work with frontend components (not required for the Python backend)

Option A — Quick start with uv (recommended by the project docs):

1. Clone the repository and navigate into it:

git clone https://github.com/keta1930/mcp-agent-graph.git
cd mcp-agent-graph

2. Initialize the MCP server using uv (Python-based approach):

uv sync
cd mag
uv run python main.py

3. Access the running application at:

• Backend API: http://localhost:9999
• API Docs: http://localhost:9999/docs
• Health: http://localhost:9999/health

Option B — Deploy with pip (traditional Python environment):

1. Ensure you are at the project root, then install dependencies and run:

cd ..
pip install -r requirements.txt
cd mag
python main.py

Option C — Docker-based deployment (using provided docker services):

1. Start Docker services defined in docker/mag_services:

cd docker/mag_services
cp .env.example .env  # edit as needed
docker-compose up -d

2. The services expose MongoDB Express at http://localhost:8081 and MinIO Console at http://localhost:9011. Backend and front-end would be accessible per the configuration in docker-compose setup.

Notes:

• The README mentions the project is paused and the online demo is discontinued, so use local deployment for exploration.
• If you modify frontend code, a separate step is available to install and run the frontend development server (optional).

Tips and common considerations:

• The project status indicates pause and discontinued online demo; expect limited maintenance and plan local testing accordingly.
• Ensure environment variables are properly configured (DB hosts/credentials, MinIO endpoints, and any API keys) when connecting to external services.
• When using Docker, verify that Docker Compose is version 2.x+ and that required ports (8081, 9011, 9999, etc.) are not in use by other services.
• The MCP integration relies on standardized adapters/tools; consult the documentation in docs/core-components and docs/first-steps/install.md for detailed usage patterns.