MolAgent

An AI agent for molecular simulation

MolAgent prepares, runs, checks and analyses simulation studies on your HPC cluster or workstation. It applies procedures written by computational chemists and works with the simulation codes your group already uses.

In alpha; private beta in 2027. The workspace and the molecular viewer are open source.

MolAgent workspace in a browser with three panes: a periodic box of a lithium-ion electrolyte in the 3D viewer, the LAMMPS input file for its equilibration, and a terminal where Packmol has built the box, next to the project file tree.
The workspace in a browser. A lithium-ion electrolyte box in the viewer, its LAMMPS input, and a terminal on the same machine, where Packmol has just built the box. The files on the right are the project folder on that machine.

How a study runs

You describe the scientific objective. The agent carries the study through four stages and stops for your input on scientific choices.

  1. Plan

    Turns the objective into a calculation plan and selects the methods and level of theory.

  2. Prepare and submit

    Builds the systems, writes the input files and submits the jobs to your scheduler.

  3. Monitor and check

    Follows the running jobs, detects errors, restarts failed steps and checks the quality of the results.

  4. Analyse

    Processes the outputs, puts the results in context and proposes the next calculations.

The agent explains its choices and asks you whenever a scientific decision is yours to make. You approve, adjust or send it back.

Procedures written by computational chemists

General-purpose language models often get specific calculations wrong. A correct result depends on choosing the method, setting up the system and the input, knowing what to watch while the job runs, testing the output and recovering from a failed run.

MolAgent gives the agent this knowledge as procedures prepared and tested by computational chemists. It combines them into studies that run several codes in sequence, with a check after each step.

Rules for shared clusters

  • No heavy computation on login nodes. The agent is blocked or asks first.
  • No software installation without your permission.
  • Working scripts are reused as they are.
  • Results are reported with their status and checked against the data before conclusions are drawn.
  • Quantum chemistry

    Gaussian, ORCA, Q-Chem, Psi4, xtb

  • Periodic DFT

    VASP, Quantum ESPRESSO, CP2K, CASTEP

  • Molecular dynamics

    GROMACS, LAMMPS, OpenMM, AMBER

  • Free energies and enhanced sampling

    PLUMED, alchemical and umbrella-sampling methods

  • Coarse-grained models

    Martini, DPD, HOOMD-blue

  • Machine-learned potentials

    MACE, CHGNet, UMA, DeePMD-kit

  • Force fields

    GAFF2, OPLS-AA, CHARMM, OpenFF

  • Building and analysis

    RDKit, ASE, pymatgen, Packmol, MDAnalysis

Capabilities: typical studies, software and checks

Mol Viewer

Structures and trajectories from the project folder open in a 3D viewer built on the Mol* engine. The viewer below is the same component, running in this page with five example systems.

Periodic box of a lithium-ion battery electrolyte: ethylene carbonate and dimethyl carbonate molecules with lithium and hexafluorophosphate ions, in ball-and-stick.

1 M LiPF6 in EC:DMC (1:1 by volume): 24 ion pairs, 180 EC and 142 DMC molecules, 3,696 atoms packed in a 34.2 Å periodic box.

Drag to rotate, scroll or pinch to zoom. The 3D viewer is about 1 MB.

The five examples: questions, systems and study plans

How it connects

A daemon runs under your user account on a cluster login node, a workstation or a server. You open the workspace in a browser or use the command line. Jobs go to the cluster scheduler or run directly on a workstation.

Your browser or terminal

The MolAgent workspace in a browser, or the command line on the machine itself.

Relay optional

For machines that accept no incoming connections. The daemon connects out to the relay, which forwards encrypted traffic it cannot read.

Your cluster or workstation

Daemon

Agent sessions, terminals and project files, under your user account.

Agent

Simulation procedures, rules for shared clusters and result checks.

Your software and data

Scheduler, simulation codes and project folders, used where they are.

Model provider

Your account with a provider, or an open-weight model on your own server. It receives the conversation and the file excerpts the agent reads.

Platform details

Your own model account

MolAgent works with the main model providers, including Anthropic, OpenAI, Google, Mistral, Amazon Bedrock, Azure and OpenRouter. Connect with an API key, or sign in with an existing subscription where the provider allows its use in third-party tools.

For confidential industrial work, an open-weight model can run on your own server or on a dedicated one, so that no external provider is involved.

Settings dialog listing model providers such as Amazon Bedrock, Anthropic, Azure and DeepSeek, with buttons to log in with an API key or a subscription.
Model providers in the workspace settings.

Open source

MolAgent publishes its workspace and its viewer under open licences and builds on open-source software. We want the simulation community to use, check and extend them.

Workspace

Daemon, relay, browser application and command line. AGPL-3.0.

Runs the agent on your machine and gives you sessions, terminals, files and viewers in a browser.

Source on GitHub

Mol Viewer

3D viewer for molecules, crystals and trajectories. MIT licence.

Viewer and light structure editor on the Mol* engine, also usable as a React component in other web applications.

@molviewer/core on npm

pi

Agent runtime. MIT licence, by Earendil.

The open-source agent harness that MolAgent runs on: tools, sessions and model providers.

pi.dev

Early access

MolAgent is in alpha. We are installing it with a small number of research groups now; a private beta opens in 2027. Write to us with the systems you study, the codes you use and the machines you compute on.

Or write to contact@molagent.ai.