ChemCrow

Status: assess | Maturity: emerging | Last reviewed: 2026-03-20

What it does

ChemCrow is an open-source LLM-based chemistry assistant that integrates GPT-4 with a suite of chemistry-specific tools. Rather than relying on the LLM’s parametric knowledge alone, ChemCrow equips the model with tool calls to cheminformatics databases and calculators — including RDKit, PubChem, and safety data sources — to answer chemistry questions, assist with synthesis planning, and reason about molecular properties.

The system was developed at EPFL and published in 2023. It represents an early instance of the “tool-augmented LLM” paradigm applied to domain science.

Key capabilities

• LLM-guided synthesis planning with tool-grounded chemical reasoning
• Integration with RDKit, PubChem, safety databases (e.g. explosive/controlled substance checks)
• Molecular property prediction (via integrated tools)
• Question answering over chemical literature and databases
• Agentic design: the LLM decides which tools to invoke at each step
• Open-source and extensible — researchers can add custom tools

Fit for Sheffield researchers

Potentially valuable for research groups in Chemistry, Chemical Engineering, and Materials Science at Sheffield. The agentic framework is a useful pedagogical example of how domain-specific tools can augment general LLMs.

However, the tool requires an OpenAI API key (GPT-4), which raises data governance considerations — research data should not be sent to OpenAI’s API without reviewing the API data handling policy. Local/open-weight LLM backends are a research direction but not yet mainstream for this tool.

Recommended as an exploratory/assess tool pending local deployment feasibility review.

Licensing and access

MIT licence. Code available on GitHub. Requires OpenAI API key. Some integrated databases (e.g. Reaxys) may require separate institutional licences.

Trust and governance notes

Queries sent to the OpenAI API for GPT-4 responses. Researchers should not send proprietary synthesis routes, unpublished compound data, or commercially sensitive information via the API without reviewing OpenAI’s enterprise API terms. Data residency is US.

Safety checks for explosive precursors are built in — a notable governance feature — but outputs should always be reviewed by a qualified chemist.

Observed limitations

• LLM can still hallucinate synthesis steps or safety information — expert review is mandatory
• Dependent on OpenAI GPT-4; open-weight alternatives not yet well supported
• Academic prototype — not a production-grade system; maintenance uncertain
• Some integrated tools (e.g. Reaxys) require separate paid access
• Primarily focused on small molecule chemistry; polymer/materials chemistry less well supported

Related tools

alphafold · labos · ag2-autogen

Sources and evidence

• ChemCrow paper: Bran et al., 2023: https://arxiv.org/abs/2304.05376 (accessed 2026-03-20)
• ChemCrow GitHub: https://github.com/ur-whitelab/chemcrow-public (accessed 2026-03-20)