MDEngine

by ForceField Silicon

A native macOS molecular-dynamics workbench: a Metal trajectory viewer, a command-line tool, and an MCP server that lets AI agents inspect trajectories and run LAMMPS simulations as detached background jobs.

Status: pre-release. Not yet distributed; the app is ad-hoc signed (builds run on the building machine only until notarization lands).

Parts

Build

Requires macOS 14+ and Xcode command-line tools. LAMMPS (brew install lammps) is needed only for run / submit_lammps.

swift build -c release            # all three products
./scripts/make_app.sh             # assemble + install /Applications/MDEngine.app

Bundled examples

Trajectory readers are safe on in-flight dumps — a file a running simulation is still writing parses to its complete frames, so you can inspect a run mid-flight. Rows with non-finite (NaN/inf) coordinates are dropped.

Both decks are ready inputs for mdengine run and submit_lammps. Bare force-field names resolve automatically: if $LAMMPS_POTENTIALS is unset, MDEngine derives it from the LAMMPS install.

MCP server

mdengine-mcp is a dependency-free MCP stdio server, listed in the MCP Registry as com.forcefieldsilicon/mdengine (server.json in this repo). Each release ships a signed macOS MCP Bundle (mdengine-mcp-<version>-macos-arm64.mcpb): double-click it to install in Claude Desktop, or unpack it (mcpb unpack) for any other client. Built from source, register with Claude Code:

claude mcp add mdengine /path/to/.build/release/mdengine-mcp

or in Claude Desktop's claude_desktop_config.json:

{ "mcpServers": { "mdengine": { "command": "/path/to/.build/release/mdengine-mcp" } } }

Hosted GPU tier over MCP (no install)

The same jobs are reachable from any MCP client that speaks HTTP — Claude Code, Claude.ai custom connectors, Cursor, Goose — via the hosted endpoint's Streamable HTTP server:

claude mcp add --transport http mdengine-cloud https://api.forcefieldsilicon.com/mcp

Sign in when the client asks (OAuth 2.1: a consent page where you paste your API key once; the client keeps a token, the key stays with you). Scripted clients may instead send the key directly as --header "Authorization: Bearer mde_YOUR_KEY". The key comes with a prepaid credit pack (forcefieldsilicon.com/mdengine). initialize and tools/list work without signing in; tool calls without a credential return 401 with the OAuth discovery pointer, which is what makes clients offer the sign-in. Tools: account, submit_job (deck inline, ≤ 8 MB), create_job + start_job (big decks via presigned PUT), job_status, job_log, job_results, list_jobs, delete_results, cancel_job. Discovery card: https://api.forcefieldsilicon.com/.well-known/mcp/server-card.json. Server code: hosted/endpoint/mde_mcp.py.

Troubleshooting the hosted connector:

Jobs run in the deck's own directory (relative read_data paths work) and launch with -sf omp -pk omp N so the OPENMP package is actually engaged; $LAMMPS_POTENTIALS is derived from the LAMMPS install when unset.

Remote hosts (run on your own Linux / GPU box)

The job runner can execute on another machine with the same tool contract: declare hosts in ~/.mdengine/hosts.json and pass host to submit_lammps (or set a default). The deck's directory is rsynced up (trajectories, checkpoints and logs excluded), LAMMPS starts under nohup with its exit code recorded remotely, and job_status / job_log / job_files / cancel_job work unchanged; fetch_job brings results back for trajectory_info, z_profile and the renderers.

{
  "default": "gpu1",
  "hosts": {
    "gpu1": {
      "ssh": "me@gpu1.example.net",
      "workdir": "~/mdengine-jobs",
      "lmp": "/usr/local/bin/lmp",
      "threads": 8,
      "launch": "{lmp} -in {input} -k on g 1 -sf kk -pk kokkos newton on neigh half -log {log}"
    }
  }
}

ssh is anything ssh accepts (key-based, non-interactive); lmp must be an absolute path (login PATH is not available over ssh); launch is optional — the default is the OpenMP form, and a GPU host is simply one whose template carries the KOKKOS flags. A remote deck must be self-contained within its own directory. The same trust model as local runs applies, on the remote machine.

Platform & limits

macOS 14+ (Apple Silicon or Intel). Trajectories are loaded whole into memory — files over 2 GB are refused with guidance to decimate or split first.

Security note

Running a LAMMPS input executes whatever the deck says. LAMMPS decks are programs, not data — they can invoke arbitrary shell commands (LAMMPS has a literal shell command). Treat a deck from someone else exactly like a shell script: read it before running it.

This applies doubly to the MCP server: an AI agent connected to mdengine-mcp can submit decks, and a submitted deck runs with your user's full privileges on this machine. That is the same trust model as any local dev tool (an agent that can run make can run anything), but be deliberate about which decks — and which agents — you hand to the job runner. Sandboxed execution (containers, no network, resource caps) is how a future hosted tier makes running untrusted decks safe; the local server does not sandbox.

Privacy

The local tools collect nothing; see PRIVACY.md, which also covers the hosted GPU tier.

License

MIT © Gitinama Inc. (d/b/a ForceField Silicon)