Nomad: Connecting AI Agents to SciFMs

Nomad is a lightweight runtime for serving scientific foundation models (SciFMs) to AI agents.

It focuses on two things:

• Serve models at scale. Run N PyTorch models across M GPUs through a unified inference interface. Nomad handles loading, scheduling, batching, and packaging, whether your weights live on Hugging Face, local storage, or in Git/Git LFS.

• Code-mode execution. Nomad exposes MCP tools as typed Python callables inside a sandbox, so agents can call and compose them efficiently within an existing execution environment.

Built on the Model Context Protocol (MCP), Nomad is agent-agnostic and designed for both HPC and local inference. It makes SciFMs easier to package, deploy, and integrate into real agent workflows.

Start Here

• Connect to a running server: Getting Started shows how to launch nomad serve, connect from MCP Inspector or Ursa, and run scripts with nomad code-mode-exec.

• Host a new model: Model Builder explains how to package a nomad.fm_base_tool.TorchModuleTool, point Nomad at model weights, and validate the resulting MCP tool.

• Operate a deployment: Deployment Guide and Deployment Specification cover the container workflow and deployment requirements.

• Work on Nomad itself: Developer Documentation collects contributor commands, docs tasks, and environment notes.

Key References

• CLI Reference

• Configuration Guide

• Configuration Reference

• Code-Mode Gateway Reference

• Tool Discovery and Model Cards Reference

• Tool Manager Reference

• TorchModuleTool Reference

• Well Format Reference

Core Concepts

• Nomad server: nomad serve loads regular Python tools and PyTorch-backed model tools from config and exposes them over MCP.

• Code-mode gateway: nomad code-mode proxies MCP servers into a sandbox where tools can be imported from mcp_tools.<server>.

• Artifact sources: model weights can come from local paths, Hugging Face, Git/Git LFS repositories, or ORAS artifacts.

• Scientific state exchange: nomad.well_format.WellFormat defines a common schema for gridded scientific state across Python objects, JSON/MCP payloads, and HDF5 files.