fairchem-mcp

An agent-steerable MCP server for FAIRChem and ASE simulations.

Most LLM-driven simulation today is batch: the agent writes a script, runs it, waits, and reads the output. fairchem-mcp makes it interactive. The model is loaded once and kept resident; relaxations and MD run in the background; and the agent can watch a simulation as it runs and steer it mid-flight — switch the optimizer when it stalls, tighten fmax, change temperature, pause, or abort.

Why

• Resident model. Load a UMA/eSEN model once; reuse it across every call. No reloading the model (seconds–minutes) on each run.

• Live monitoring. get_status returns step, energy, max force, and a trend verdict (decreasing / plateaued / stuck / diverging) so the agent can decide whether to intervene.

• Mid-flight steering. steer can pause / resume / abort, set_fmax, switch_optimizer, or set_temperature on a running job. Switching optimizer carries the atomic positions over — it just rebuilds the driver.

• Hybrid namespace. High-level tools (start_relaxation, …) and the execute / inspect_expr escape hatch share one Python namespace, so the agent can drop to raw Python on the very same live Atoms.

• Code awareness. introspect reads the installed API (real signatures and docstrings) and live objects via jedi — not possibly-stale docs.

Related MCP server: MOF Tools MCP Server

Install

pip install -e .              # core: mcp + ase + jedi + numpy (works with EMT)
pip install -e ".[fairchem]"  # add FAIRChem (torch + models) for UMA/eSEN
pip install -e ".[lammps]"    # add LAMMPS as a classical force engine
pip install -e ".[saddles]"   # add Sella + POUNCE for transition-state searches

ASE, numpy, jedi and mcp are core dependencies (installed automatically) — the server is fully usable out of the box with the built-in EMT calculator.

FAIRChem (optional)

pip install -e ".[fairchem]" pulls in fairchem-core (PyTorch + models). To actually load a UMA model you also need:

• a Hugging Face account with approved access to the facebook/UMA model repository, and huggingface-cli login (or HF_TOKEN) set;

• PyTorch for your platform (CUDA build for GPU; CPU works but is slower).

Everything except load_model works with a plain ASE calculator (attach_emt), so you can develop and test without a GPU or model.

LAMMPS (optional)

pip install -e ".[lammps]" installs the lammps Python package. LAMMPS is used as a force engine (classical potentials) while ASE drives the dynamics, so attach_lammps works with every steerable job (MD, relaxation, NEB, phonons, EOS, minima search).

macOS note (Homebrew MPICH): the PyPI lammps wheel links @rpath/libmpi.12.dylib and libpmpi.12.dylib, which the dynamic loader can't find by default. If you have Homebrew's mpich (brew install mpich), attach_lammps auto-symlinks those libs into the lammps package directory on first use. If that can't be applied (read-only install, non-Homebrew MPI), either symlink them yourself or export before launching the server:

export DYLD_FALLBACK_LIBRARY_PATH=/opt/homebrew/lib:$DYLD_FALLBACK_LIBRARY_PATH

Verify with: python -c "from lammps import lammps; lammps(cmdargs=['-log','none','-screen','none']).close(); print('ok')".

Transition-state searches (optional)

pip install -e ".[saddles]" adds two extra saddle-point backends: sella (a rational-function ASE optimizer) and pounce-solver (multistart eigenvector following). The dimer method (start_saddle_search) is pure ASE and needs neither. See examples/saddles/ for all three.

Register with Claude Code

Add to your MCP config (see examples/claude_mcp_config.json):

{
  "mcpServers": {
    "fairchem": { "command": "fairchem-mcp" }
  }
}

Tools

Resources: sim://models, sim://job/{id}/status, sim://job/{id}/trajectory.

Example flow

attach_emt()                      -> calc_1
build_structure({"kind":"bulk","name":"Cu","crystalstructure":"fcc",
                 "a":3.6,"repeat":[2,2,2],"rattle":0.2})  -> struct_1
start_relaxation("struct_1","calc_1",optimizer="FIRE",fmax=0.01)  -> job_1
get_status("job_1")               -> {status:"running", trend:{label:"stuck", ...}}
steer("job_1","switch_optimizer",optimizer="LBFGS")
get_status("job_1")               -> {status:"converged", ...}
introspect("atoms", live=True)    -> live object signature/docstring

Finding multiple relaxed geometries

start_minima_search finds several distinct local minima of the PES — useful for surface adsorption sites, cluster isomers, or conformers. It relaxes repeatedly from the starting structure on a PES biased to repel the minima already found, then polishes each escape on the true PES:

• kernel="flooding" (default) adds Gaussian bumps (sigma Å, amplitude eV); kernel="deflation" adds inverse-distance poles (eta, power). Both are best for fixed-frame problems (an adsorbate on a frozen slab, an anchored conformer).

• kernel="basinhopping" (random kick + relax + Metropolis accept) is the right tool for free clusters / nanoparticles, whose rigid-body rotation defeats a spatial bias. Pair it with comparator="fingerprint" (see below).

• New minima are deduplicated by energy (energy_tol) plus a structure comparator: "rmsd" (raw coords, frame-dependent — fine for a fixed frame) or "fingerprint" (sorted pairwise distances; rotation/translation/permutation invariant — use for free clusters and molecules, or rotated copies get miscounted as distinct). Each accepted minimum is registered as its own structure.

This reuses the escape mechanism from POUNCE's find_minima (deflation / flooding) but drives it with ASE's gradient optimizers — the right inner solver for a PES — so each escape relaxation is a normal steerable job (watch the trend, switch_optimizer, set_fmax, pause/abort). POUNCE's interior-point solver is deliberately not used as the inner relaxer.

Examples

examples/catalysis/ has four end-to-end catalysis workflows — adsorption energy, adsorption-site search, diffusion-barrier NEB, and surface-stability phonons — each as a runnable script and an MCP tool-call walkthrough. They run on EMT out of the box; set FAIRCHEM_MCP_EXAMPLE_MODEL for UMA.

examples/saddles/ covers the three single-ended transition-state routes — dimer, Sella, and POUNCE eigenvector following — on one shared system so you can compare them.

examples/alloys/ builds up to alloy design: the elastic stiffness tensor (start_elastic_scan), the formation-energy convex hull (start_convex_hull), and a design loop that gates candidates on stability then ranks the survivors by a mechanical property.

Safety

execute / inspect_expr run arbitrary Python in-process. This is a trusted local developer tool — do not expose it to untrusted input or over a network.

Notes

• Only one job runs at a time (serializes model/GPU access).

• All optimizers/integrators use logfile=None; the stdio transport reserves stdout for the MCP protocol.