ChemCrow

What is ChemCrow?

ChemCrow is an open-source package for solving reasoning-intensive chemistry tasks with a language model and specialist tools. Built with LangChain, it connects model reasoning to RDKit, PaperQA, chemical databases such as PubChem and ChemSpace, and services for retrosynthetic planning and reaction-product prediction. The package is released under the MIT license.

ChemCrow can answer quantitative molecular questions, search chemical information, reason over research literature, and apply reaction or synthesis tools. The public repository notes that it does not include every tool described in the ChemCrow research paper because some APIs have usage restrictions, so the package should not be expected to reproduce all paper results.

By default, some reaction capabilities depend on the RXN4Chem API and require credentials. Developers can optionally self-host prepared Docker images for reaction prediction and retrosynthesis and run ChemCrow with local reaction services. Model use also requires the developer to supply a supported OpenAI API credential, while SerpAPI is optional for search.

How to Use ChemCrow

1. Install the package with pip install chemcrow.
2. Set the required model API credential and, optionally, a SerpAPI credential.
3. Import ChemCrow, select the model and temperature, and create the chemistry agent.
4. Submit a chemistry question or task through the agent's run method.
5. If local reaction tools are needed, start the provided Docker images and enable local reaction mode.
6. Review all generated chemical reasoning and proposed reactions with qualified human oversight.

Core Features

• Chemistry tool use: Connects language-model reasoning with molecular and reaction software.
• Molecular calculations: Uses RDKit and related utilities for chemical properties and structures.
• Literature retrieval: Applies PaperQA to questions grounded in scientific publications.
• Chemical databases: Queries sources such as PubChem and ChemSpace.
• Reaction planning: Supports retrosynthesis and reaction-product prediction through remote or local services.
• Open-source package: Provides inspectable Python code under the MIT license.

Use Cases

• Chemical question answering: Researchers calculate or retrieve molecular information.
• Literature-supported analysis: Scientists investigate questions using relevant papers.
• Synthesis planning research: Chemists explore potential retrosynthetic routes with tool assistance.
• Reaction prediction: Developers test likely products for specified reactants.
• Agent research: AI researchers study how language models behave when equipped with chemistry tools.

Pricing

ChemCrow itself is open source under the MIT license. Running it can still incur charges from the selected language model, search service, or reaction-planning APIs.

Frequently Asked Questions

Does the public package reproduce every ChemCrow paper result?

No. The repository explicitly states that some paper tools are absent because of API restrictions.

Can reaction tools be self-hosted?

Yes. The repository provides Docker commands for local prediction and retrosynthesis services.

Is ChemCrow a substitute for professional chemical review?

No. Its outputs are generated research assistance and should be evaluated by qualified users before real-world decisions or experiments.