Every result is a file with a download URL rather than a blob in the chat, which is what makes a trajectory usable at all. For interfaces, doped structures or custom slabs the builder stops short — generate those with ASE or pymatgen yourself and hand them back through write_structure_workflow.
A FastMCP server over ASE, pymatgen and machine-learned interatomic potentials. Results are written as files and returned as download URLs, so binary output never lands in the conversation.
- Structures built as bulk, surface, molecule, supercell, amorphous, liquid, bicrystal or polycrystal — build_structure_workflow
- Geometry optimisation with convergence, step and constraint controls — optimize_structure_workflow
- Energy, forces and stress at a fixed geometry, without modifying the structure — single_point_workflow
- Molecular dynamics under Velocity Verlet, Langevin or Berendsen thermostats — run_md_workflow
- Analysis: RDF and coordination statistics, MSD and thermodynamic trends, VACF and Green-Kubo diffusion — analyze_structure_workflow, analyze_trajectory_workflow, autocorrelation_workflow
- A geometry persisted for a later step, including structures built elsewhere — write_structure_workflow
- Artifacts downloadable as xyz, extxyz, cif, traj, png, svg, csv and dat
Python 3.11 or newer. Install with pip install -r requirements.txt, then serve Streamable HTTP with uvicorn mcp_atomictoolkit.http_app:app, or run python -m mcp_atomictoolkit.mcp_server for stdio. The default potential is kim; nequix and orb are also supported. A Dockerfile and a render.yaml are included.