Changelog

What is new and what changed on Cheméo, most recent first.

2026-08-03

MCP server changes

• An estimate declared out of range no longer carries an error bar. When Cheméo says it cannot estimate the uncertainty on a molecule far outside a model's experience, that number is now gone from the whole answer rather than only from the summary line. It had remained a level down, where a reader looking for it would have found it and believed it.
• Every answer names the version that produced it, with a pointer to this page. Cheméo changes under long-running work, and there was no way to tell from an answer that anything had moved since the last one.
• The similarity score says what it means. Estimates from Cheméo Relay come with how closely the molecule resembles what the model was trained on, and that number reads as a promise about accuracy, which it is not. It says how well the training data covers the structure; the uncertainty says how accurate the estimate is. The answer now says so, once, beside the method.
• Citations point at the paper. Where a method's record carries the DOI of the work it comes from, that is now the link given for it rather than a page describing the method.
• A measured zero is reported. A property whose value is exactly zero came back with the value missing while the counts beside it still said a measurement was there. Propane's dipole moment is zero, because the molecule is symmetric, and that is the expected answer for every symmetric compound rather than an edge case. 80 stored properties were affected. Where Cheméo deliberately reports no representative value, as when measurements straddle zero and their middle would mean nothing, it says so as it always has.
• An out-of-domain estimate reports how far out it is, even at zero. A similarity of zero to everything a model was trained on is the strongest warning available, and it was the one case where the number went missing.

2026-08-02

Changed

• Corrected the Joback group decomposition (v23). Cheméo estimates several properties with the Joback group-contribution method (Joback and Reid, 1987), which breaks a molecule into functional groups and sums a contribution for each. Two groups in that table were defined by the wrong structural pattern: the pattern named "ester" in fact matched a peroxide, and the one named "carboxylic acid" a hydroperoxide. Any molecule carrying an ester or a carboxylic acid was therefore split into the wrong fragments, and every Joback estimate derived from that split was wrong. Aspirin, which carries both groups, had a Joback critical pressure of 4,083 kPa; it is now 3,577 kPa.

• Aligned the Joback boiling-point constant to the 198.2 published in the original paper, from the 198.0 Cheméo was using.

• Which estimates changed. All of these are Joback: ideal gas heat capacity on 72,850 compounds, liquid dynamic viscosity on 36,042, critical pressure on 22,053, enthalpy of fusion on 38,999, and standard Gibbs energy of formation on 2,959. The two curves changed on every compound that has them because Cheméo plots them between Joback endpoints, heat capacity from the boiling point to the critical temperature and viscosity from the melting point to the boiling point, so the corrected boiling point shifts the whole temperature range each is computed over.

• What did not change. No experimental measurement was touched. Cheméo Relay predictions are identical, and Relay is what Cheméo displays for boiling point, melting point and critical temperature, so those are unaffected on the compound page even though Joback's own estimate of them moved. The Wildman-Crippen logP and the McGowan characteristic volume are also unchanged: neither uses Joback. v21 and v22 were internal development databases.

MCP server changes

• A predicted heat capacity now carries its temperature. Properties that vary with temperature or pressure came back as bare numbers, with nothing to say what conditions they applied to. A heat capacity without a temperature is not data. Each predicted point now states its conditions, and each property states the range it covers, the same way measured values always have.
• Predictions say when they are extrapolations. Asked about a molecule unlike anything the model has seen, Cheméo used to answer with the same confidence it brings to a familiar one. It now says so, and how far outside its experience the molecule sits. The number is still returned: when nothing better exists you may well want it, and you can weigh it once you know.
• Predictions state their uncertainty and are rounded to it. A critical temperature whose model error is 33 K is no longer reported to six figures.
• A new list_properties tool. The property keys were not guessable, and the acentric factor being af rather than omega was enough to send a caller down the wrong path. Every key is now listed in a single call that costs no quota, with its unit and whether it is one value or a table.
• A property with no meaningful summary no longer invents one. Some reaction properties pool different reactions, so their values run from strongly negative to strongly positive. Where that happens no representative value is reported, and the reason is stated; the individual values and their range are still there.
• Smaller prediction answers, which leaves more room for the work you are actually doing.
• Better suggestions when a search goes nowhere. A SMILES Cheméo cannot read now points you at searching by name or CAS number instead, and a search that finds nothing points at prediction, but only when what you typed really was a structure.
• The tool now shows what the website shows. Where Cheméo holds both a measurement and an estimate for the same property, the compound page has always shown the measurement alone, and where it holds two estimates it shows the released model rather than the beta one. The tool was showing all of them and reporting the middle value, so a boiling point could come back as the average of a measurement and a neural network estimate, labelled experimental. On 42% of compounds the reported number moves. It now matches the page.

2026-08-01

Changed

• Search results are ranked by relevance. Searching for a compound by name now returns that compound first. Before this, results came back in Cheméo id order, so searching "ethanol" gave 757 matches with Ethanol on none of the first pages. The order is deliberately simple and readable: an exact name or synonym match, then a name that starts with what you typed, then one that contains it as a word, with better characterised compounds first within each.
• An exact match is called out. When what you typed is exactly a compound's name or one of its synonyms, that compound is marked in the results table, even among thousands of candidates. It is the difference between "this scored highest" and "this is the compound you named".
• Every result says how much data Cheméo holds for it. A new Data column counts the properties on each compound, so you can tell at a glance which of several candidates is worth opening.
• Searching by structure and by formula is exact again. Case carries meaning in chemical notation, and Cheméo was losing it: Co (cobalt) and CO (carbon monoxide) were treated as the same thing, as were the aromatic and aliphatic forms of a SMILES, so a benzene structure search could return cyclohexane. They are now distinct. Formulas Cheméo previously did not recognise as formulas, including AgCl, SbCl5 and SiF4, are recognised, and you can search by InChI.

MCP server changes

• The beta caveat about ranking is gone, because the reason for it is. The search tool reports ordering: "relevance", carries the exact match as its own field, and no longer advises searching by CAS or SMILES to work around the result order.
• Two new search options: exact, which returns only a whole-name match and returns nothing rather than falling back to a fuzzy search, and has_cas, which drops compounds Cheméo holds no CAS number for.

2026-07-30

New

• An MCP server, in beta. Cheméo now speaks the Model Context Protocol, so an AI assistant can look up property data directly instead of reading the website. Connect it at https://www.chemeo.com/api/mcp. You sign in and approve the connection yourself, the access is read only, and you can disconnect any application at any time from your profile.
• Four tools. Search for compounds, read the full property set for up to 25 of them at once, ask where a specific value came from, and run a property prediction from a SMILES string.
• Answers you can check. Every value says whether it is an experimental measurement or a prediction, how many measurements and how many independent sources back it, the spread between them, and how many were excluded as outliers. Every result carries the citation, and where a value names an original publication it says to cite that ahead of Cheméo. A property Cheméo does not hold is stated outright rather than left as a gap.
• Beta means the tools and their answers may still change. (Search results are ranked by relevance as of 2026-08-01; when this was written they were not.)

2026-07-26

Added

• 4,672 recommended liquid heat-capacity points across 376 compounds (v19). 3,279 isobaric liquid heat capacity and 1,393 saturation heat capacity. Each is a critically-reviewed recommended value that carries the source citation and the monograph table reference.
• Source: Zabransky, M., Ruzicka, V., Majer, V. and Domalski, E. (1997), Heat Capacity of Liquids: Critical Review and Recommended Values, Monograph No. 6, J. Phys. & Chem. Ref. Data (JPCRD), NIST, Gaithersburg, MD.

2026-07-19

New

• A bigger database (v18). Cheméo now covers more than 103,000 compounds and 17,400 mixtures, about 940 compounds and 1,200 mixtures more than before, from rebuilding the whole database against the current source data.
• More measured data. Thousands of added experimental values, especially phase-transition points: over 6,000 more boiling points and nearly 3,000 more melting points, plus more critical temperatures and enthalpies of vaporization. New temperature and pressure dependent property types are included too, such as saturation heat capacity and isothermal compressibility.

Changed

• Predicted values you can rely on. A predicted property is now shown only when an experimentally measured compound sits near it in chemical structure space. Predictions that fall outside the models' validated range are no longer shown as numbers. Instead the compound page points you to the prediction tool, which still computes them on demand. The aim is fewer numbers, and more trustworthy ones.

Improved

• Stereoisomers stay distinct. cis/trans and alpha/beta forms are kept as separate compounds instead of being merged into a single entry.
• More structures kept. Compounds whose structure our chemistry toolkit cannot build (some organometallics, ions, and noble-gas dimers) are now retained with their measured data rather than dropped.

2026-07-16

New

• Search for mixtures by their components. A search like "methanol water" now finds the mixtures that contain both, alongside the pure compounds. Component synonyms are matched too, so "methanol" finds mixtures listed under "Methyl Alcohol". Mixtures appear in their own section on the results page, each linking to its mixture page.

Improved

• Faster "similar compounds". The list of similar compounds on every compound page is computed noticeably faster.
• Worklist checks entries. Adding a compound or mixture to a worklist now confirms it exists first, and each list has a size limit.
• Style and script updates apply right away. After an update you no longer need to force-refresh the page to get the latest styling or behaviour.

2026-07-15

Changed

• Updated property database (v17). Predicted values are now stored next to the experimental data for the same property, so parity plots are generated near instantly in the prediction tool. Displayed averages and ranges continue to use the experimental data.

API

• Request rate limits. The public API now allows up to 120 requests per minute, with short bursts of up to 30, and reports the current limit in standard RateLimit response headers so clients can pace themselves.

2026-07-14

Changed

• A refreshed look. Cheméo has a cleaner, modern design: refreshed styling throughout, clearer and easier-to-read property tables, crisp icons, and a layout that makes better use of wide screens.