MCP Server (AI Agents)

Foundry includes an MCP (Model Context Protocol) server that enables AI assistants like Claude to discover and use materials science datasets.

What is MCP?

MCP is a protocol that allows AI assistants to use external tools. With Foundry's MCP server, you can ask Claude:

• "Find me a materials science dataset for band gap prediction"

• "What fields are in the OQMD dataset?"

• "Load the training data and show me the first few rows"

Quick Start

Install for Claude Code

foundry mcp install

This adds Foundry to your Claude Code configuration. Restart Claude Code to activate.

Manual Start

For custom integrations:

foundry mcp start

Available Tools

The MCP server provides these tools to AI agents:

search_datasets

Search for materials science datasets.

Parameters:

• query (string, required) - Search terms

• limit (integer, optional) - Maximum results (default: 10)

Returns: List of datasets with name, title, DOI, description

Example prompt: "Search for datasets about crystal structures"

get_dataset_schema

Get the schema of a dataset - what fields it contains.

Parameters:

• doi (string, required) - The dataset DOI

Returns: Schema with splits, fields, data types, and descriptions

Example prompt: "What fields are in dataset 10.18126/abc123?"

load_dataset

Load a dataset and return its data with schema.

Parameters:

• doi (string, required) - The dataset DOI

• split (string, optional) - Which split to load (default: all)

Returns: Data with schema information and citation

Example prompt: "Load the training data from the band gap dataset"

list_all_datasets

List all available Foundry datasets.

Parameters:

• limit (integer, optional) - Maximum results (default: 100)

Returns: Complete catalog of available datasets

Example prompt: "What datasets are available in Foundry?"

Configuration

Claude Code

The foundry mcp install command adds this to your Claude Code config:

{
  "mcpServers": {
    "foundry-ml": {
      "command": "foundry",
      "args": ["mcp", "start"]
    }
  }
}

Custom Integration

For other MCP-compatible clients, the server uses stdio transport:

from foundry.mcp.server import create_server

config = create_server()
# config contains server name, version, and tool definitions

Example Conversations

Finding a Dataset

You: Find me a dataset for predicting band gaps of materials

Claude: I'll search for band gap datasets in Foundry.

Uses search_datasets tool

I found 5 datasets related to band gaps:

1. OQMD Band Gaps (10.18126/abc) - 50,000 materials

2. AFLOW Band Gaps (10.18126/def) - 30,000 materials ...

Understanding a Dataset

You: What's in the OQMD band gaps dataset?

Claude: Let me get the schema for that dataset.

Uses get_dataset_schema tool

The OQMD Band Gaps dataset contains:

• Inputs: composition (chemical formula), structure (crystal structure)

• Targets: band_gap (eV)

• Splits: train (80%), test (20%)

Loading Data

You: Load the training data and show me some examples

Claude: I'll load the training split.

Uses load_dataset tool

Here are the first 5 rows:

composition	band_gap
Si	1.12
GaAs	1.42
...

Troubleshooting

Server Not Starting

Ensure Foundry is installed correctly:

pip install --upgrade foundry-ml
foundry version

Tools Not Available

Restart Claude Code after installing:

foundry mcp install
# Restart Claude Code

Authentication Issues

For datasets requiring authentication, ensure you've authenticated:

from foundry import Foundry
f = Foundry()  # This triggers auth flow if needed

Technical Details

Protocol

The MCP server uses JSON-RPC 2.0 over stdio.

Server Info

from foundry.mcp.server import create_server

config = create_server()
print(config)
# {
#   "name": "foundry-ml",
#   "version": "1.1.0",
#   "tools": [...]
# }

Tool Definitions

Each tool follows the MCP tool schema:

{
    "name": "search_datasets",
    "description": "Search for materials science datasets...",
    "inputSchema": {
        "type": "object",
        "properties": {
            "query": {"type": "string", "description": "..."},
            "limit": {"type": "integer", "default": 10}
        },
        "required": ["query"]
    }
}