openmm-md

OpenMM MD (free, Windows-native GPU molecular dynamics)

openmm-md is a FREE / open-source (MIT) molecular-dynamics MCP that runs natively on Windows with CUDA on the local NVIDIA GPU (verified on an RTX 4050). It is the free lane for MD work; it is explicitly a lesser alternative to Schrodinger Desmond, which stays the preferred engine whenever a legitimate license is available.

When to use which engine (decision guide)

Situation	Use	Why
On Windows, no Schrodinger license, need MD now	openmm-md (this)	Only native-Windows GPU MD option; free; runs on the local RTX 4050
Quick stability / RMSD / Rg / H-bond check of a pose	openmm-md	Fast, free, good enough for triage
Ranking a congeneric ligand series roughly, no budget	openmm-md openmm_mmgbsa	Free relative ranking (approximate)
You need validated FEP+, WaterMap, IFD-MD, Prime MM-GBSA (VSGB2.1)	Schrodinger (mcp__schrodinger__*)	OpenMM has no validated equivalent
Publication-grade / regulatory binding free energies	Desmond / FEP+ on a licensed host	More-validated force fields + workflows
You have a legit Schrodinger license + a Linux box	Desmond on Linux	The mcp__schrodinger__desmond_* tools work there
You want a big GPU / no local install	Desmond on a licensed Linux cloud GPU	Scale out with a real license

Golden rule: if the science needs a validated commercial method, use Schrodinger. Use openmm-md when free + Windows-native + "good enough" is the right trade. openmm-md does NOT replace the 67 mcp__schrodinger__* tools; they stay intact and preferred.

Honest OpenMM vs Desmond trade-offs

OpenMM (free) is a genuinely excellent MD engine, but as a workflow it is a lesser alternative to Desmond/Maestro:

• Force fields: OpenMM here uses AMBER ff14SB (protein) + OpenFF Sage or GAFF (ligand). Desmond ships OPLS4, extensively validated for drug discovery.
• Ligand charges (Windows caveat): AmberTools has no native-Windows build, so there is no antechamber / sqm AM1-BCC and no GAFF here. Charges come from NAGL (a GNN AM1-BCC surrogate, if installed) or RDKit Gasteiger (lower fidelity). Both are clearly labelled. OpenEye/sqm AM1-BCC quality is not reproduced.
• Free-energy / analysis: no FEP+, no WaterMap, no IFD-MD. openmm_mmgbsa is an HONEST OBC2/GBSA single-trajectory approximation (no entropy, generic radii) -- good for ranking, NOT absolute affinities, and materially weaker than Prime MM-GBSA.
• Polish: no Maestro GUI, trajectory clustering, or event analysis suite. You get MDTraj metrics (RMSD/RMSF/Rg/H-bonds/contacts).

Running Desmond legitimately (the better option when licensed)

This project will not install, crack, or facilitate a pirated Schrodinger suite. To use Desmond properly:

1. Get a legitimate license. Academic/non-commercial Desmond is free from D. E. Shaw Research (deshawresearch.com) for non-commercial use; commercial use needs a Schrodinger license. Academic Maestro/Schrodinger licenses are available to universities.
2. Desmond on Linux. Desmond's MD engine is Linux-only. Install the licensed suite on a Linux workstation, set $SCHRODINGER, and the existing mcp__schrodinger__desmond_setup / desmond_md / desmond_trajectory_analysis tools drive it there.
3. Desmond on a cloud GPU. Put a legit-licensed Linux install on a cloud GPU instance (e.g. via the cloud-hub RunPod / Modal / Lightning GPU tools), copy the license, and run Desmond remotely. Shape only -- provision the box, install YOUR licensed suite, run.

The openmm-md workflow

Tools chain through a job_dir on disk. Typical protein-ligand run:

1. openmm_status -- verify CUDA is live on the GPU (run this first).
2. openmm_prepare -- protein = PDB path or 4-char RCSB id; optional ligand_file (a docked SDF/MOL2 pose -- preferred) or ligand_smiles (arbitrary conformer). PDBFixer cleanup + ff14SB + OpenFF ligand params. Returns the job_dir.
3. openmm_solvate -- TIP3P box (box_shape cube|dodecahedron|octahedron) + neutralizing ions (PME). hydrogen_mass_amu=4.0 for HMR / 4 fs.
4. openmm_minimize -- energy minimization on CUDA.
5. openmm_equilibrate -- NVT then NPT (barostat), optional heavy-atom restraints.
6. openmm_production -- production MD. Non-blocking: returns a handle; poll with openmm_production(job_dir, poll=True) or Read prod_status.json. Writes prod.dcd. Use blocking=True only for short runs.
7. openmm_analyze -- MDTraj RMSD / RMSF / Rg / H-bonds / protein-ligand contacts + ligand binding-site RMSD (stability).
8. openmm_mmgbsa -- approximate free binding dG (ranking only; read fidelity_note).

Sensible defaults / tips

• GPU memory (6 GB RTX 4050): keep systems modest (small-to-medium proteins, ~1.0 nm padding). Large systems may OOM -- reduce padding or trim the system.
• Timestep: 2 fs default (HBond constraints). 4 fs only with HMR (hydrogen_mass_amu=4.0).
• Ligand pose: for binding studies pass a docked pose via ligand_file (e.g. from Glide mcp__schrodinger__glide_dock); a bare SMILES has no bound pose.
• Long runs: leave production non-blocking and poll; the detached worker keeps writing prod.dcd + prod_status.json even across MCP calls.
• Honesty: label OpenMM results as free-stack / approximate; when a decision hinges on absolute affinity or validated methods, escalate to Schrodinger.

Registration

MCP server: C:\Users\Prekzursil\source\repos\openmm-md-mcp\server.py, run with the conda env Python D:\micromamba\envs\openmm-md\python.exe (the server self-configures the Windows CUDA DLL path, so no activation wrapper is needed). See the repo README for the exact ~/.mcp.json block and how to install this skill into a skills directory.