One Sentence, One Full Quantum Portrait

We typed one sentence — “characterize caffeine” — and got back a complete quantum portrait, without touching a single input deck. No basis-set syntax, no keyword blocks, no queue script. Just a plain-language request to the Hilbeon assistant, which drove the engine over its MCP / HTTP interface and reported real numbers. Every figure below is that run — nothing is illustrative.

This is what “agentic quantum chemistry” actually means in practice: the tool speaks chemistry, so you don't have to speak Fortran. Here is the thread, lightly trimmed.

you    ▸ characterize caffeine

Hilbeon
  ✓ built C8H10N4O2 from SMILES (24 atoms, RDKit ETKDG+MMFF)
  ✓ RHF/6-31G(d)    E = -676.3278 Ha   (216 basis functions)
  ✓ B3LYP/6-31G(d)  E = -679.9798 Ha
  · HOMO -5.83 eV · LUMO -0.71 eV · gap 5.13 eV  (B3LYP)
  · dipole 3.56 D   → matches the measured ~3.6 D
  · frontier orbital rendered below

That is the entire interaction. What used to be an afternoon of deck-writing, basis lookups, and unit conversions collapses into a request and a rendered answer. Let's read what the electrons said.

The molecule, built and solved

Caffeine's SMILES becomes a 24-atom, C₈H₁₀N₄O₂ structure, and the engine converges a real self-consistent-field wavefunction over 216 basis functions. Two levels of theory ran back to back: Hartree–Fock for the reference energy, and B3LYP for the physically meaningful frontier gap. The interactive orbital below is the actual computed HOMO — the same cube file the engine wrote, streamed live.

Does the number match reality?

The test of a calculation is not that it is precise — it is that it is right. Caffeine has a large, well-measured dipole moment: experiments put it in the 3.58–3.83 D range depending on solvent. Hilbeon's computed dipole is 3.56 D — sitting right at the lower edge of that band, exactly where a gas-phase, single-conformer calculation should land relative to solution measurements that polarize the molecule upward. No fitting, no tuning: the electrons simply reproduce the molecule's known polarity. The frontier gap of 5.13 eV is likewise physically sensible for a colorless, UV-absorbing aromatic. The portrait isn't just internally consistent; it agrees with the bench.

The receipts

Property	Hilbeon computed	Reference / note
Structure from SMILES	C₈H₁₀N₄O₂ · 24 atoms	RDKit ETKDG + MMFF, single point
Total energy — RHF/6-31G(d)	−676.3278 Ha	reference wavefunction, 216 basis functions
Total energy — B3LYP/6-31G(d)	−679.9798 Ha	same geometry; physical level for the gap
HOMO / LUMO / gap — B3LYP	−5.83 / −0.71 eV; gap 5.13 eV	the physically meaningful frontier gap
HOMO / LUMO / gap — RHF (Koopmans)	−8.53 / +2.98 eV; gap 11.51 eV	HF overestimates gaps — expected, shown for honesty
Dipole moment	3.56 D	experiment ~3.58–3.83 D (solvent-dependent)

The RHF frontier gap is in the table on purpose: Hartree–Fock inflates gaps, so we report the B3LYP value as the physical one and show the Koopmans number beside it rather than quietly dropping it. That is the house style — the honest number, with its caveat, every time.

Honest edges

Three caveats travel with this run, and the assistant states them rather than hiding them. The geometry is an RDKit MMFF conformer with single-point energies on top, not a QM-optimized structure — good for a first portrait, worth upgrading before you quote a strain energy. The frontier orbitals are Kohn–Sham / Koopmans-level reads, excellent for reasoning about reactivity trends, not a substitute for a computed ionization potential. And the dipole is one conformer in vacuum. None of that weakens the point: the workflow that produced these numbers was a single plain-language request, and the headline number lands on the measured one. (For what it's worth, every energy the engine reports is also checked against an independent reference code before it ships — quietly, as internal QA. The number we put in front of you here is the one that matches the lab.)

From one sentence to your pipeline

Caffeine is a friendly first subject, but the mechanism is the product. Ask in words; the assistant plans the calculation, runs the owned integral engine, and hands back a portrait you can act on. When you outgrow the chat, every step is already an API call. That is the whole arc — from a SMILES string, or even just a name, to a full quantum portrait, in one thread.

References

• Caffeine overview — PubChem: Caffeine
• Caffeine dipole moment (experimental range ~3.58–3.83 D) — Sci. Rep. (caffeine, methods & properties)
• Model Context Protocol — modelcontextprotocol.io