Model Context Protocol (MCP) Integration

This document describes how to set up and use Model Context Protocol (MCP) with Aurelian agents.

What is MCP?

MCP (Model Context Protocol) is an open protocol that standardizes how applications provide context to LLMs. It works like a "USB-C port for AI applications," providing a standardized way to connect AI models to different data sources and tools.

Key benefits:

Connect LLMs to pre-built integrations with a standard protocol
Flexibility to switch between LLM providers
Secure your data within your infrastructure
Enable complex agent workflows

Setting Up MCP Servers for Aurelian

Aurelian provides MCP server implementations for several of its agents. You can use these servers to expose Aurelian's capabilities to any MCP-compatible client (such as Claude).

Configuration Options

You can configure MCP servers in two ways:

Manual Configuration: Create the full JSON configuration directly
Generated Configuration: Use the provided config generator script

Option 1: Manual Configuration

Create a file named mcp-config.json with your server configurations:

{
  "mcpServers": {
    "memory": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-memory"
      ],
      "env": {
        "MEMORY_FILE_PATH": "~/.mcp/memory.json"
      }
    },
    "linkml": {
      "command": "python",
      "args": [
        "/path/to/aurelian/src/aurelian/agents/linkml/linkml_mcp.py"
      ],
      "env": {
        "AURELIAN_WORKDIR": "/tmp/linkml"
      }
    }
  }
}

Option 2: Generated Configuration

Create a simplified configuration file (e.g., simple-config.json):

{
  "memory": {
    "type": "memory",
    "memory_path": "~/.mcp/memory.json"
  },
  "linkml": {
    "type": "linkml",
    "workdir": "/tmp/linkml",
    "python_path": "/usr/bin/python"
  }
}

Generate the full configuration:

python src/aurelian/mcp/config_generator.py --config simple-config.json --output mcp-config.json --base-dir /path/to/aurelian

Available Agent Servers

Aurelian provides the following MCP servers:

Server Name	Type	Description
linkml	Agent	LinkML schema validation and data modeling
gocam	Agent	Gene Ontology Causal Activity Models
phenopackets	Agent	Working with phenotype data in Phenopacket format
robot	Agent	Ontology manipulation with ROBOT
amigo	Agent	Gene Ontology and gene associations
uniprot	Agent	UniProt protein information
diagnosis	Agent	Disease diagnosis with Monarch Knowledge Graph
memory	Utility	Persistent memory for MCP interactions

Configuration Parameters

Common Parameters for Agent Servers

type: The agent type (e.g., "linkml", "gocam")
workdir: Directory for the agent to store files (default: /tmp/{agent_type})
python_path: Path to Python executable (default: /usr/bin/python)
email: Email address for services that require it (optional)
doi_urls: DOI URL pattern for literature-related agents (optional)
env: Additional environment variables (optional)

Memory Server Parameters

type: Always "memory"
memory_path: Path to store memory file (default: ~/.mcp/memory.json)

Custom Server Parameters

type: "custom"
command: Executable command
args: List of command arguments
env: Environment variables dictionary

Running MCP with Claude

Install the MCP CLI:

npm install -g @modelcontextprotocol/cli

Start the MCP servers:

mcp start --config mcp-config.json

Open Claude and ensure MCP is enabled in settings
In Claude, you can now interact with your Aurelian agents through MCP

Example Use Cases

LinkML Validation: Use Claude to validate data against LinkML schemas
Ontology Management: Create and manipulate ontologies through natural language
Gene Association Analysis: Ask questions about gene functions and associations
Phenotype Analysis: Work with structured phenotype data

Troubleshooting

Ensure the correct Python environment is activated for each server
Check server logs for errors by running mcp logs
Verify paths in your configuration are correct and accessible
Ensure required environment variables are set properly