Chemistry AI workflows often combine molecular data, scientific software, model predictions, and explanations from an assistant. Reusing one requires understanding what each step means, which inputs it accepts, and what evidence supports its output. A saved conversation or a script can be useful, but neither automatically communicates all of those decisions to another researcher.
CARS (Chemistry AI Recipe Specification) is an open, vendor-neutral community draft initiated by ChemAI Atlas to explore a shared description of chemistry Recipes. As checked on 11 October 2026, its status is Community Draft 0.1. This guide explains the proposal, reads an actual repository example, and separates available structural validation from future execution and scientific evaluation.
Why chemistry workflows are hard to reuse
Consider a researcher who receives a molecular structure, calculates selected descriptors, and shares a results table. Another researcher needs more than the column names: Which chemical representation was accepted? Were mixtures rejected? Was stereochemistry retained? Which method and software version produced the values? Are the numbers calculations, predictions, or measurements?
These are illustrative questions for reviewing a workflow, not results from a calculation performed for this article. Their value is that they expose decisions that can otherwise remain inside code, notebooks, or a particular application. A reusable description should make those decisions inspectable before someone chooses an execution environment.
The CARS draft proposes describing purpose, chemical inputs and outputs, capabilities, limitations, security, and evidence together. That can help authors discuss a workflow even when they use different tools. Whether an eventual implementation preserves those meanings across systems still needs demonstration; the description alone does not establish interoperability. See the draft object model.
What is CARS?
A CARS Recipe is proposed as an inspectable manifest: a structured document explaining a chemistry task and its contracts. The draft separates the scientific purpose from optional execution bindings. A descriptive Recipe can therefore communicate intent without claiming that a machine can run it.
CARS is not a hosted chemistry AI model, a universal workflow engine, or an adopted industry standard. The project README calls the format experimental and its fields provisional. “Vendor-neutral” expresses the design goal of avoiding dependence on one platform; it is not evidence that independent implementations already exchange Recipes successfully.
ChemAI Atlas provides the project overview and a Recipe catalog, while GitHub is the authoritative location for the evolving CARS draft. A catalog Recipe is not automatically CARS-compliant or CARS-certified. The word Recipe describes a useful workflow concept in both contexts, but compliance would require a declared specification version, applicable checks, and evidence.
Define → Validate → Connect
The overview uses Define → Validate → Connect as a conceptual lifecycle. It is a way to organize responsibilities, rather than an executable pipeline available today.
Define: describe the task and its boundaries
The proposed manifest records the Recipe identity and version, purpose, inputs, outputs, capabilities and dependencies, scientific context, and security declarations. For chemistry, specifying a file format is only a beginning: structure representation, mixture policy, units, and result meaning also matter.
The draft distinguishes deterministic calculations, model predictions, retrieved evidence, human annotations, experimental measurements, and generated explanations. This distinction helps readers ask what kind of evidence a value actually represents. Naming a result kind does not prove that the producer used the declared method correctly. The specification describes these proposed semantics.
Validate: check the document before trusting the result
The repository contains a provisional JSON Schema and a reference Python validator. Its documented scope is L0 structural validation only: it checks document shape and the restrictions encoded in that schema. It does not inspect molecular chemistry, resolve an adapter, audit permissions, execute a calculation, or assess scientific evidence.
The schema notes explicitly describe a subset of the prose proposal. A structurally valid document is consequently a starting point for further review, rather than a certificate of safety or accuracy.
Connect: explore execution bindings
The draft discusses optional bindings to CWL, Arazzo, Agent Skills, MCP, or native implementations. These are proposed connection mechanisms. A reference to a workflow file is not the file itself, and naming an adapter does not demonstrate an installed, tested implementation.
The execution interoperability RFC asks which mechanisms should be prioritized. Anyone evaluating an integration should request its supported mapping, dependencies, environment, and run evidence. This article does not claim that those integrations exist.
Read a real molecular-descriptor example
The repository currently includes molecular-descriptors.yaml. The following is a verbatim excerpt from its output declaration, not a complete manifest:
outputs:
- id: "descriptors"
type: "molecular_properties"
formats: ["application/json"]
properties:
- name: "molecular_weight"
unit: "g/mol"
result_kind: "deterministic_calculation"
id names the output, type states its chemistry-oriented category, and formats declares JSON. The property identifies molecular weight, gives its unit, and classifies it as a deterministic calculation. This is a declaration of expected meaning; it contains no computed molecular-weight value.
Elsewhere, the example requires a chemical structure in SMILES format, rejects mixtures, and asks that stereochemistry be preserved if present. It declares structure validation before descriptor calculation and limits applicability to single-component organic molecules supported by the selected toolkit. Its scientific validation status is not_assessed. Its illustrative input is CCO, with numerical output deliberately omitted pending an executable test.
The example also names a CWL binding to workflow.cwl. The schema notes say this reference is intentionally non-executable and must be replaced before claiming execution support. No molecular calculation was run for this guide. The excerpt is attributed to ChemAI Atlas's CARS repository under its declared CC BY 4.0 recipe-document license; omitted fields remain available in the complete source.
How CARS relates to MCP, CWL, and Agent Skills
These technologies address different responsibilities. The comparison describes their roles, not verified CARS integrations.
| Technology | Primary role | Boundary relevant to CARS |
|---|---|---|
| CARS, Community Draft 0.1 | Proposed chemistry Recipe descriptions and scientific context | An experimental description does not supply a runtime or scientific certification |
| MCP | Connect AI applications with external tools and data sources | Tool connectivity does not by itself establish chemical meaning or workflow validity |
| CWL | Describe command-line tools and composed workflows, including execution semantics | A CARS binding still needs a real CWL document and an appropriate runner |
| Agent Skills | Package agent instructions with optional scripts, references, and resources | Instructions still depend on the host, tools, environment, and permissions |
The role definitions come from the MCP introduction, CWL standards, and Agent Skills overview. CARS aims to complement these mechanisms with chemistry-focused descriptions, rather than replace them.
OpenAPI describes HTTP API interfaces, while Arazzo describes API call sequences and dependencies. They are relevant when a workflow calls services. RO-Crate packages research data with structured metadata; it addresses a different documentation need. None of these roles demonstrates a working CARS adapter. For broader vocabulary, see the existing MCP, Skills, agents, and APIs guide.
Scientific validation goes beyond a successful run
Three questions must remain separate: Does the manifest match the schema? Does an implementation execute successfully under declared conditions? Does the result support the intended scientific conclusion? A positive answer to one leaves the others open.
For a descriptor calculation, a proposed review would examine input identity, normalization decisions, units, method and version, and appropriate reference values. For a predictive workflow, it would additionally examine evaluation data, possible training overlap, uncertainty, applicability domain, and failures. These are editorial review questions consistent with the draft's evidence goals, not a report of completed benchmarking or experimental confirmation. The draft validation framework discusses task-specific scientific evaluation.
Even the validation labels remain under discussion. The scientific validation RFC proposes L0 schema, L1 semantics, L2 execution, and L3 scientific evidence review. The inspected specification instead proposes L0 schema, L1 semantics, L2 security, L3 execution, and L4 scientific assessment. Readers should cite the document and revision with a level number. Neither proposal defines finalized certification; the implemented reference validator covers L0 only.
Likewise, declared evidence labels do not establish evidence on their own. Benchmarking, independent review, and experimental support answer different questions. For practical documentation principles, continue with reproducibility for chemistry AI workflows.
Current status and limits
At the inspected repository revision 6c2049c, the available artifacts include the draft specification, one example manifest, the provisional schema, validator source, automated test definitions, and a GitHub Actions workflow configured to validate the schema and examples and run tests. Reading those definitions does not establish that a particular CI run passed. This article claims no validator run, end-to-end execution, or scientific validation of the example.
The roadmap describes planned prototype deliverables. It is not a record of completed milestones. Broader semantic, security, execution, and scientific checks remain future work according to the schema notes. The contribution guide also says formal governance remains to be agreed.
Use the draft now to inspect terminology and propose clearer contracts. Before relying on execution portability, request implementation evidence. Before relying on a scientific result, review task-specific evidence. The present sources do not establish external adoption, standards-body endorsement, or CARS certification of catalog Recipes.
How to participate
Chemists can contribute concrete cases: ambiguous structure handling, a missing unit, an unsuitable output type, or evidence needed to interpret a prediction. Developers can explain how those requirements would map to a real tool or runtime, including failure behavior and compatibility limits.
Start with the current specification. Join the discussions on chemistry-specific data types, scientific validation, or execution interoperability. Use Issues for actionable defects or proposals, and follow the contribution guide when submitting example or documentation improvements through pull requests.
CARS offers a community space to make chemistry workflow intent and evidence easier to inspect. Its usefulness must develop through concrete examples, critical review, and demonstrated implementations. Read the draft and join GitHub Discussions to help shape that work.
Further reading and source scope
Sources were checked on 11 October 2026. CARS implementation claims refer to repository revision 6c2049ccf10f22c0cd5ab361711bbc8e65bb02e7; links to main invite participation in the evolving draft. Specification text and documentation are CC BY 4.0. This guide paraphrases the proposal and explains a short attributed excerpt.
Explore RDKit and the existing Molecular Analysis with Claude + RDKit MCP Recipe for related chemistry tooling and a catalog workflow. These links provide context; they do not claim CARS compatibility. The CARS overview and repository remain the starting points for the initiative.