mcp-atomictoolkit

An MCP-compatible server providing atomistic simulation capabilities through ASE, pymatgen, etc.

⚛️ MCP Atomic Toolkit

Python 3.11+ License: MIT MCP tests

[!NOTE] This project is under active development. Interfaces and behavior may evolve.

A FastMCP server for atomistic modeling workflows powered by ASE, pymatgen, and modern ML interatomic potentials.

It gives MCP clients a practical toolkit for:

  • building structures,
  • running geometry optimization + molecular dynamics,
  • analyzing structures/trajectories,
  • and downloading generated artifacts (data + plots).

✨ Why this repo

If you need atomistic workflows exposed as MCP tools (instead of hand-wiring scripts), this project gives you:

  • ready-to-call MCP tools for common simulation tasks,
  • file-first outputs that are easy to inspect/reuse,
  • artifact download URLs so clients don’t need binary blobs in chat context,
  • deployment-ready HTTP app with health and server-card endpoints.

🚀 Features

  • MCP-native workflows via FastMCP tools
  • Structure generation: bulk, surface, molecule, supercell, amorphous, liquid, bicrystal, polycrystal
  • Optimization workflows with MLIPs (kim default, nequix/orb supported)
  • Molecular dynamics workflows (Velocity Verlet, Langevin, NVT Berendsen)
  • Analysis outputs:
    • RDF + coordination stats
    • MSD + thermodynamic trends
    • VACF + diffusion (Green-Kubo)
  • Downloadable artifacts (xyz, extxyz, cif, traj, png, svg, csv, dat, ...)
  • Registry-friendly endpoints (/healthz, server card, Streamable HTTP root)

⚡ Quick Start

1) Requirements

  • Python 3.11+

2) Install

pip install -r requirements.txt

3) Run locally

uvicorn mcp_atomictoolkit.http_app:app --host 0.0.0.0 --port 10000

Alternative:

python main.py

STDIO mode (for desktop MCP clients):

python -m mcp_atomictoolkit.mcp_server

[!IMPORTANT] STDIO transports must keep stdout clean for JSON-RPC. Avoid print() or logging to stdout when running the server in STDIO mode.

4) Smoke check

curl -s http://localhost:10000/healthz

Expected response:

{"status":"ok"}

🧰 Tooling Overview

Main MCP tools exposed by the server:

  • build_structure_workflow
  • analyze_structure_workflow
  • write_structure_workflow
  • optimize_structure_workflow
  • single_point_workflow
  • run_md_workflow
  • analyze_trajectory_workflow
  • autocorrelation_workflow

Legacy aliases are also included for backward compatibility.


🌐 Endpoints

  • POST / — primary MCP Streamable HTTP endpoint
  • GET /healthz — health check
  • GET /docs — lightweight documentation (README)
  • GET /.well-known/mcp/server-card.json — MCP server card metadata
  • GET /artifacts/{artifact_id}/{filename} — artifact download route
  • /sse/ — compatibility alias path mounted to the MCP app

📦 Deployment

Render

render.yaml is included and ready to use.

Default start command:

uvicorn mcp_atomictoolkit.http_app:app --host 0.0.0.0 --port $PORT

Docker

docker build -t mcp-atomictoolkit .
docker run --rm -p 7860:7860 mcp-atomictoolkit

🗂️ Project Structure

src/mcp_atomictoolkit/
  mcp_server.py          # FastMCP tool definitions
  http_app.py            # Starlette app + routing/endpoints
  workflows/core.py      # High-level workflow orchestration
  analysis/              # Structure/trajectory/VACF analysis logic
  structure_operations.py
  optimizers.py
  md_runner.py
  artifact_store.py      # Download artifact registration + URLs

🧪 Workflow Notes (for MCP clients)

Structure building coverage

build_structure_workflow supports:

  • bulk (ASE bulk)
  • surface (ASE surface)
  • molecule (ASE molecule)
  • supercell (multiplication of a base structure)
  • amorphous/liquid (random packed structures)
  • bicrystal and polycrystal (grain stacking/rotation)

For interfaces, doped structures, adsorbates, or custom slabs, prefer:

  1. Generate the structure with ASE/pymatgen (or an external builder), then
  2. Use write_structure_workflow to persist the final geometry for downstream steps.

This ensures MCP callers can still handle advanced structures even when a specialized builder is required.

Builder kwargs cheat sheet

Common builder_kwargs for build_structure_workflow:

  • surface: indices, layers, vacuum
  • supercell: size, base_structure_type, base_crystal_system, base_lattice_constant, base_kwargs
  • amorphous/liquid: num_atoms, box_length, relax, relax_steps, relax_fmax
  • bicrystal: grain_size, interface_axis, rotation_angle, rotation_axis, interface_gap
  • polycrystal: num_grains, grain_size, rotation_angle

Optimization options

optimize_structure_workflow exposes:

  • max_steps, fmax (convergence)
  • maxstep, alpha (BFGS step/damping controls)
  • constraints (fixed_atoms, fixed_bonds, fixed_cell)

Single-point calculations

single_point_workflow computes energy, forces, and stress (if periodic) without modifying the structure, making it suitable for quick evaluations.

MD integrators / ensembles

run_md_workflow supports:

  • velocityverlet / nve (NVE)
  • langevin / nvt-langevin (NVT)
  • nvt / nvt-berendsen (NVT)

Tune temperature_K, friction, and taut to control thermostat behavior.


📈 GitHub Pulse

Add your repository path in the URLs below to enable live charts.

Star history

Star History Chart


🤝 Contributing

  • Keep outputs file-based and artifact-friendly.
  • When adding tools, usually update both:
    • workflows/core.py
    • mcp_server.py
  • Preserve http_app.py compatibility behavior unless intentionally changing deployment contracts.

📄 License

MIT — see LICENSE.

Máy chủ liên quan

NotebookLM Web Importer

Nhập trang web và video YouTube vào NotebookLM chỉ với một cú nhấp. Được tin dùng bởi hơn 200.000 người dùng.

Cài đặt tiện ích Chrome