Skip to content

ChemGraph

Thang Pham, Murat Keçeli, Aditya Tanikanti / Argonne National Laboratory

ChemGraph is a Python agent framework that connects natural-language chemistry requests to molecular construction, simulations, analysis, and reporting, with CLI, Python, Streamlit, and MCP interfaces.

Catalog updated ·

Website ↗GitHub / Repository ↗Documentation ↗
Follow

Public link：https://chemaiatlas.com/agents/chemgraph/

Claim this resource

Overview

ChemGraph provides an orchestration layer for computational chemistry and materials-science workflows rather than a standalone predictive model. Its source-described architecture combines LangGraph with ASE, RDKit, and MCP to translate natural-language requests into tool-assisted work. Researchers can access it through a command-line interface, an asynchronous Python API, a source-tree Streamlit interface, or an MCP tool server. The repository supplies runnable code; its package metadata labels the software as beta.

Inputs include chemistry questions, requested calculation settings, and, for specialized workflows, ligand/receptor information or PDF/text documents. Outputs can include molecular structures, calculation results, trajectories, spectra, and reports, stored in session artifact directories. The default single_agent workflow is the recommended entry point. Other workflows support planner/executor decomposition, document retrieval, docking, XANES tasks, and interactive workspace work. The checkpointed main_agent can discover instruction-based Skills and delegate to configured specialists; those Skills are supporting instructions, not separate autonomous agents.

Available calculations depend on detected engines and installed dependencies. EMT and MACE are included in the core installation, while other calculators and specialized workflows require additional packages, executables, or credentials. EMT is described as suitable for setup checks, not general high-accuracy chemistry. Distributed execution requires site-specific configuration and infrastructure. ChemGraph can launch calculations and modify files, so generated inputs, units, convergence, and conclusions require scientific review. Workspace shell access is not confined to the workspace, despite action-review mechanisms. The official README documents these capabilities and boundaries; they are not evidence of independent scientific validation.

Key Features

  • Connects natural-language requests to molecule lookup, molecular construction, ASE calculations, analysis, and report generation.
  • Offers CLI and asynchronous Python access, a Streamlit interface requiring a source checkout, and MCP tool serving over stdio or streamable HTTP.
  • Provides a default `single_agent` workflow alongside specialized docking, document-retrieval, XANES, and planner/executor workflows with documented dependency requirements.
  • Supports durable interactive `main_agent` sessions, on-demand tool discovery, instruction-based Skills, optional specialist delegation, and action reviews for file mutations and shell commands.
  • Detects available calculator engines at startup and exposes those present in the environment, with EMT and MACE included in the core installation.
  • Includes optional local-process, Parsl, Ensemble Launcher, Globus Compute, and Academy execution paths that require additional dependencies or infrastructure.

Use Cases

  • Suggested evaluation: compare a natural-language molecule lookup or structure-building request against a manually checked chemical identifier and structure.
  • Suggested evaluation: run a small ASE calculation with an explicitly selected calculator, then inspect generated inputs, energies, units, and convergence before considering broader research use.
  • Suggested evaluation: combine PDF/text retrieval with chemistry tools to assess whether answers remain traceable to the supplied documents.
  • Suggested integration evaluation: expose ChemGraph chemistry tools to an MCP client, or assess a distributed screening workflow after configuring the required execution backend and site resources.

How to Use

  1. Read the installation guide and prepare an isolated Python environment. The supplied README requires Python 3.11 or newer; install only the optional dependencies needed for your chosen workflow.
  2. Select an LLM provider using Models and authentication. Configure the required credentials outside committed files, and distinguish provider access from the locally installed ChemGraph framework.
  3. Follow the quickstart with single_agent. Begin with a molecule lookup or an explicitly selected EMT setup check; lookup examples may require network access, and EMT is not a general high-accuracy chemistry method.
  4. Inspect the session artifacts, normally written under cg_logs/. Check structures, calculator settings, units, convergence, and final interpretations before relying on results.
  5. Consult the workflow guide before adding docking, retrieval, XANES, or workspace tasks. Review approval behavior and host-access boundaries for interactive agents.
  6. For tool integration, follow the MCP server guide. Treat distributed execution as a separate setup task requiring its documented dependencies and infrastructure.

Related resources

Allegro

Model

Allegro implements an E(3)-equivariant interatomic potential as a NequIP extension, with documented GPU acceleration options and a separate plugin for LAMMPS simulations.

Open sourcePython

Computational Chemistry · Materials Discovery

Updated

ASE Skill (Jinzhe Zeng Group)

Skill

An ASE routing Skill in the computational-chemistry-agent-skills collection that separates workflow preparation from calculator configuration and delegates execution elsewhere.

Computational Chemistry · Materials Discovery

Updated

ChemAgent (AI4Chem)

Agent

ChemAgent (AI4Chem) is a research framework for chemistry and materials tool use, linked to the CheMatAgent paper on tree-search planning, tool execution and ChemToolBench-based training.

Computational Chemistry · Materials Discovery

Updated

CHGNet

Model

CHGNet is a pretrained, charge-informed neural network potential for crystal structures, predicting energies, forces, stresses and magnetic moments for relaxation and molecular dynamics workflows.

Open sourcePython

Computational Chemistry · Materials Discovery

Updated

DeePMD-kit

Model

DeePMD-kit is a toolkit for training and fine-tuning Deep Potential interatomic models from quantum-mechanical reference data, then exporting them for inference and molecular dynamics.

Open sourcePythonC++

Computational Chemistry · Materials Discovery

Updated

DP-GEN

Open Source

DP-GEN is a Python concurrent-learning platform that coordinates molecular simulation, first-principles calculations and DeePMD-kit workflows to generate interatomic potential models.

Open sourcePython

Computational Chemistry · Materials Discovery

Updated

Related guides

Agents

AI Agents for Chemistry: From Chatbots to Autonomous Research

A practical guide to eight chemistry and biomedical research agents: how tools, memory, planning and multi-agent roles work, which projects fit different tasks, and how to evaluate bounded autonomy with scientific oversight.

· 10 min read
Overview

MCP vs Skills vs Agents vs APIs: What's the Difference?

Understand how APIs, MCP servers, Skills, and agents divide interface, procedural guidance, execution, and planning responsibilities in chemistry workflows—and how to choose or combine them without confusing access with scientific validity.

· 10 min read
Overview

How to Build a Chemistry AI Agent Stack

Design a chemistry AI agent stack by separating language models, orchestration, Skills, MCP interfaces, APIs, libraries and data. Follow proposed molecular and materials workflows, define input/output contracts, and plan permissions, evaluation and reproducible deployment.

· 9 min read