Crystallography MCP server

CIF inspection, d-spacings, powder patterns, and triacylglycerol polymorph fingerprinting, backed by pymatgen.

Built 21 August 2026 for the Laurelate bay-nut-fat work, but domain-general.

Why this and not a VESTA MCP

VESTA has no scripting interface — no documented command-line arguments, no batch mode, no macros, no headless operation. An MCP over it could only run open -a VESTA file.cif and screenshot the window. pymatgen exposes the actual science, so that is what this wraps.

Keep VESTA as a human viewer. It is genuinely good at that.

Related MCP server: mofstructure-mcp

Setup

Needs Python ≥3.10 (current pymatgen requires it) and its own venv:

python3 -m venv ~/.venvs/crystal
~/.venvs/crystal/bin/python -m pip install -r requirements.txt

Register with Claude Code at user scope, so it loads in every session:

claude mcp add -s user crystallography -- \
  "$HOME/.venvs/crystal/bin/python" \
  "$HOME/Documents/crystallography-mcp/server.py"

Check and remove:

claude mcp list
claude mcp remove -s user crystallography

It works with any MCP client, not just Claude Code — it is a plain stdio server, so point your client at <venv python> server.py.

Tools

The multi-datablock trap

Some deposited CIFs hold more than one structure. vanlangevelde2000_CLC_MPM_bprime.cif contains both CLC (a = 57.368 Å, the β′ template that matters) and MPM (a = 76.21 Å). Taking block 0 silently is how you end up analysing the wrong structure. list_datablocks exists to make that visible.

Polymorph fingerprints

Diagnostic subcell short spacings, Å:

Validated against both solved templates, which have known answers:

• vanlangevelde2000_CLC_MPM_bprime.cif block 0 → BETA_PRIME ✓

• vanlangevelde1999_PPP_beta.cif → BETA ✓

And against synthetic patterns built from each reference line set → correct verdict in all three cases.

Known limits

• Intensities are not Rietveld-grade. Tabulated form factors, Debye–Waller omitted. Peak positions are exact; relative intensities are good enough to rank lines, not to refine occupancies.

• Peak picking is deliberately simple — local maxima with a flat-baseline prominence filter. For a weak or noisy scan, lower min_prominence and eyeball the picked list before trusting the verdict.

• spglib may disagree with the CIF's declared space group. It re-derives symmetry from coordinates, and published long-chain structures often carry disordered or split chain-end sites. Loosen symprec before concluding anything.

• compare_structures returning False across different chain lengths is expected, not a bug — a C16 template will never match a C10/C12 fat.

Provenance note

This server found a real bug in R&D/modeling/baynutfat/structure.py: the I-centring translation was being applied twice for β′-CLC (the CIF already lists all 8 centred operators), leaving 4 duplicated carbons — 284 C where there should be 280. Peak positions were unaffected, structure factors were slightly off. Guarded now by acceptance test T10 and modeling/tests/crosscheck_pymatgen.py.

That is the argument for having a second implementation available.

Citing

If a result from this ends up in a publication, cite pymatgen — it does the real work:

S.P. Ong et al., Comput. Mater. Sci. 68, 314–319 (2013).

And if you rendered anything in VESTA alongside it, VESTA asks to be cited too:

K. Momma and F. Izumi, J. Appl. Crystallogr. 44, 1272–1276 (2011).

License

MIT — see LICENSE.