ChemCrow
Open-source LLM agent that plans and runs chemistry tasks with real tools.
About
ChemCrow (Bran et al., Nature Machine Intelligence 2024) augments an LLM with 18 expert-designed chemistry tools — molecule lookup, reaction prediction, retrosynthesis planning, safety checks, and literature search — wired through a ReAct-style reasoning loop. It planned syntheses of an insect repellent and organocatalysts and helped discover a novel chromophore.
Notably, it includes a built-in safety check that halts when an input is identified as a controlled chemical. The code is open source (pip install chemcrow) but needs an OpenAI API key and, for full function, access to reaction-prediction services. It is a research prototype, not a validated lab assistant — treat its plans as drafts for a chemist to review.
What it can do
- Retrosynthesis planning
- Reaction product prediction
- Molecule property lookup
- Literature-grounded answers via paper search
- Controlled-chemical safety refusal
Setup
- Install with: pip install chemcrow
- Set OPENAI_API_KEY in your environment (and SERP_API_KEY for web search, optional).
- Create an agent: from chemcrow.agents import ChemCrow; chem_model = ChemCrow(model="gpt-4-0613", temp=0.1).
- Run tasks: chem_model.run("What is the molecular weight of tylenol?")
- For full retrosynthesis tools, self-host the provided Docker images or use the RXN4Chem API.
Prerequisites
Python, OpenAI API key; GPU optional for self-hosted reaction tools.
Use cases
Works with
Trust score — 7/10
Scored 7/10 (verified tier, metrics-v1). Evidence: influential paper (Nature Machine Intelligence 2024) but niche practitioner base; research prototype; community forks keep it alive; fully open source; pip-installable. Manual review: important proof of concept for chemistry agents, including the built-in controlled-chemical safety check — but verify every plan with a chemist. License: Open source; see repo for license..
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 6.0/10
- Maintenance · recency, releases, issue load
- 6.0/10
- Openness · license + docs
- 10.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · influential paper (Nature Machine Intelligence 2024) but niche practitioner base
- · research prototype; community forks keep it alive
- · fully open source; pip-installable
- · manual review: important proof of concept for chemistry agents, including the built-in controlled-chemical safety check — but verify every plan with a chemist
Sources
Every fact above was checked against these sources. Where a claim comes from the provider and couldn't be independently verified, the text says so.