Organic Chemistry · Mechanism Explorer
Sean's three templates — Robinson annulation, Grignard reagents, Diels–Alder — are now all generated by the app from whatever starting materials you type, alongside SN2. Nothing is looked up. Aromatic rings are read, and the product's configuration is drawn with wedges where the chemistry sets it.
Type the starting materials and the app reads them, finds the reactive sites, decides which pathway those sites support, and derives every step by applying curly arrows to the previous structure. The intermediates are what the arrows produce, never drawings from memory, so a substrate nobody prepared still works — and a combination that does not react is refused with the chemistry reason.
Development status shows what is live, what was last verified, what is next and every change, regenerated on each publish.
Try Sean's own examples: 1,3-cyclohexanedione with MVK using NaOH · bromobenzene with Mg, then CO2 and H3O+ · ethyl acetate with 2 equivalents of ethylmagnesium bromide · Diels-Alder of cyclopentadiene with acrolein · SN2 of 2-bromobutane with cyanide
Any ketone or 1,3-dione with an enolizable alpha position, with any Michael acceptor that has an alpha-prime hydrogen. Eight elementary steps plus the product, with every proton transfer as its own panel, exactly as Sean marked up — the Michael addition is steps 1–3 and the aldol condensation steps 4–8, the same split as his template. The base is whatever the request names: NaOH in water, as he draws it, or NaOEt, NaOMe, KOtBu.
Any primary or secondary alkyl chloride, bromide or iodide with any nucleophile with a lone pair. A neutral nucleophile gets the extra proton transfer it needs. If the carbon is a stereocentre the first and last panels carry wedges and the product is drawn with the inverted configuration — the umbrella turning inside out.
Any R–MgX with an aldehyde, ketone, ester, epoxide or carbon dioxide; or an organohalide plus magnesium, in which case the formation of the reagent is the first panel. An ester collapses to the ketone and takes the second equivalent, giving the tertiary alcohol; CO2 gives the carboxylic acid; an epoxide opens at its less substituted carbon. The aqueous acid workup is drawn as its own proton transfer from hydronium. A substrate with an O–H or N–H is refused: the Grignard is destroyed by it before anything else can happen. Alkyl halides are refused too — Grignards do not do SN2.
Any conjugated diene with any alkene or alkyne dienophile. One concerted step, three arrows. Which way round the partners join is decided from the donor and withdrawer positions (the ortho/para outcome, worked out from orbital coefficients rather than remembered). A ring-locked diene gives the bridged bicyclic; an alkyne gives a 1,4-cyclohexadiene; benzoquinone takes a second diene on request. The product's stereocentres are set from a model of the transition state: cis substituents stay cis, trans stay trans, and the endo product is the one drawn.
Configuration is stored on the structure the way SMILES states it and never depends on the drawing. Wedges and hashes are then chosen for each panel by embedding every candidate in three dimensions and recomputing the configuration — a wedge is never placed by convention, only by that test. Write a SMILES with @ or / and \ to specify a starting configuration; a cis dienophile is drawn cis and stays cis.
Where it is drawn: the SN2 inversion, and the four new centres of a Diels–Alder adduct (endo, with retention of the dienophile's geometry). Where it is not yet: the ring-fusion carbon of a Robinson product and the new carbon of a Grignard addition are formed as racemates and are left unwedged; a bicyclic product is drawn flat with its bridge inside the ring, so read each wedge at its own carbon rather than comparing wedges across the ring.
Benzene rings are read from names (benzaldehyde, 4-methoxybenzaldehyde, 1-bromo-4-nitrobenzene, phenylmagnesium bromide) and from aromatic SMILES, which is kekulised on the way in. The ring's alternating bonds are recognised structurally and left alone: an aryl ketone is not a Michael acceptor, a benzene ring is not a diene, bromobenzene does not do SN2 but does form a Grignard reagent.
It reproduces Sean's signed-off mechanism exactly. Generated from scratch for 2-methylcyclohexanone and MVK, the output matches his marked-up mechanism panel for panel — same atoms, bonds, charges and arrows across all nine panels. That equivalence is a test, so it stays true. His new 1,3-cyclohexanedione example produces the same nine panels he drew.
An outside valence model checks every intermediate. Every derived structure across 34 generated mechanisms — 185 panels covering all four families and the aromatic cases — is rebuilt in RDKit and sanitised there. Nothing was rejected.
330 mechanisms across the substrate space are generated and inspected as drawings — bond lengths, overlaps, arrows anchored to things that exist — every time the tests run. The regiochemistry rules and the stereochemical outcomes (cis, trans, endo) are asserted in the same suite.
A tertiary halide for SN2, an acceptor with no alpha-prime hydrogen, a Grignard with an alcohol or acid, a benzene ring offered as a diene, phenyl vinyl ketone as a Robinson acceptor. Each refusal gives the reason, because the reason is the teaching.
ANTHROPIC_API_KEY on the Cloudflare project, on your account so the usage bills to the product. Everything above works without it.More examples of the families already built, especially the cases a student gets wrong — they become tests. And the choice of the fourth reaction: an aldol or Claisen condensation reuses most of what is here. Any generated mechanism can still be printed as a single sheet with every panel and every arrow labelled, for marking up on paper.