NYU, Matterstack ship AskChem, a 2.4M-claim chemistry MCP index

A small paper out of NYU and Matterstack points at something worth watching in the RAG-for-science space: swap the retrieval unit from paper to atomic claim, and you can apparently stop language models from making up citations. AskChem indexes 2.4M provenance-carrying claims extracted from 147K chemistry papers, each grounded by a source DOI and a verbatim quote, per the arXiv paper. It is live at askchem.org with a web UI, REST API, SDK, and MCP server, so any agent that speaks Anthropic's Model Context Protocol can hit the claim store as a tool.

The evaluation is the interesting part. On AskChem-Bench, a suite of 30 cross-paper chemistry questions, a GPT-5.5 reader grounded through AskChem returned 100% resolvable DOIs, versus 88.3% for the same reader running from parametric memory alone. The authors also report the highest citation density among five tested systems at 18.1 verified DOIs per answer, ahead of Paperclip, Edison Scientific's PaperQA-family agent, and Google NotebookLM Deep Research. A domain-expert audit of a stratified sample of 148 edges in the 171,342-edge evidence graph put edge-type precision at 97.9% across 146 decidable relations.

Why this matters if you are not building chemistry tools: the MCP delivery is the template. Verifiable-retrieval infrastructure becomes a service any research agent can subscribe to rather than a bespoke integration, and 'claim as the retrieval unit' is a design that could port to other scientific domains once someone does the extraction work.

The honest caveats are worth naming. Thirty questions is a narrow benchmark, and the authors themselves say it measures groundedness rather than full factual accuracy or user utility. Edison Scientific, a closed agentic system, produces substantially more citation-linked quantitative detail and a slightly higher on-topic rate than AskChem, so cross-paper reliability is not settled. And while 100% of the 2.4M claims carry a source DOI and quote, the paper is clear that traceability is not the same as semantic correctness. An LLM can still misinterpret what a source actually says.

The direction is the part to watch. If claim-level infrastructure with open data can match the trust properties of closed research assistants, the pricing and access picture for academic labs changes, and Matterstack has planted an early flag on what could become a common substrate for scientific agents.