FutureHouse

FutureHouse is a nonprofit artificial intelligence research organisation focused on accelerating scientific discovery through advanced AI systems. Based in San Francisco, California, the organisation develops and evaluates AI agents designed to help scientists investigate complex research questions, analyse scientific literature, generate hypotheses, and support experimental research. Its work focuses primarily on biology and other complex scientific fields, with a broader mission of using AI to advance fundamental research that can improve human health.

Founded in 2023 by Sam Rodriques and Andrew White, FutureHouse was established around the idea of building an AI Scientist: an AI system capable of carrying out important parts of the scientific research process. Traditional scientific work requires researchers to read large amounts of literature, identify unanswered questions, develop hypotheses, plan experiments, interpret results, and communicate findings. These activities can take considerable time, particularly when research involves multiple disciplines and large scientific datasets. FutureHouse develops AI systems intended to help researchers manage some of this complexity and investigate scientific questions more efficiently.

FutureHouse operates as an independent nonprofit research laboratory rather than a conventional software subscription company. Its activities combine AI research, scientific experimentation, open-source development, research evaluation, and programmes that connect AI technology with working scientists. The organisation develops and validates research systems while making selected tools and research resources available to the wider scientific community. Its work is particularly relevant to academic researchers, biomedical scientists, biotechnology teams, computational biologists, and developers building AI applications for scientific discovery.

One of FutureHouse’s important contributions is its development of specialised AI agents for scientific literature research. These agents are designed to do more than retrieve a list of publications. They can search scientific sources, examine relevant evidence, compare findings, and produce answers supported by research references. This approach is intended to make scientific information easier to investigate while allowing researchers to inspect the evidence behind an answer.

The organisation introduced the FutureHouse Platform in May 2025, making a collection of scientific AI agents available through a web interface and an API. The original platform launched with four specialised agents: Crow, Falcon, Owl, and Phoenix. Each agent was designed for a different research task, allowing scientists to select an appropriate system based on the question or workflow they wanted to investigate. The platform demonstrated how specialised agents could be combined to support more complex scientific research activities.

Crow was developed as a general-purpose scientific literature research agent. It can search scientific publications and provide concise answers to research questions. Its purpose is to help researchers find relevant information without having to manually examine every publication returned by a search. Crow was also designed for API-based workflows, allowing developers to incorporate scientific question-answering capabilities into their own research applications. This makes it relevant to teams building automated literature monitoring, scientific knowledge retrieval, and research assistance systems.

Falcon was designed for deeper literature reviews. It can search and synthesise information across a larger body of scientific publications, helping researchers investigate topics that require evidence from multiple studies. Its original platform integration included access to specialised scientific databases such as OpenTargets. Researchers can use this type of capability to explore disease mechanisms, examine relationships between biological targets, compare research findings, and identify disagreements between publications. Deep literature synthesis can also help researchers understand the strengths and limitations of existing evidence before planning further investigations.

Owl, previously known as HasAnyone, was developed to help researchers determine whether a particular scientific idea or experiment had already been investigated. Establishing whether previous work exists is an important part of research planning because scientists need to understand existing knowledge before proposing new experiments. Owl was designed to search for relevant prior work and help researchers investigate whether a proposed approach has already appeared in the literature. This can support research novelty assessments, literature reviews, and the identification of possible gaps in scientific knowledge.

Phoenix was introduced as an experimental chemistry-focused agent based on ChemCrow. It was designed to work with specialised chemistry tools and assist with research questions involving chemical compounds and experimental planning. Its original capabilities included supporting investigations into molecular properties, chemical novelty, possible reactions, and compound selection. FutureHouse identified Phoenix as experimental and noted that it had not been benchmarked as extensively as the other original agents. Its outputs therefore required particular care and scientific review.

Another major development from FutureHouse is PaperQA2, an AI-powered research tool for question answering over scientific documents. PaperQA2 uses retrieval-augmented generation to search documents, gather relevant evidence, and generate answers supported by in-text citations. It can work with PDFs, text files, web-based documents, Microsoft Office files, and other supported document formats. Researchers can use it to investigate a collection of papers, ask questions about the information they contain, and obtain answers connected to specific passages in the source material.

PaperQA2 is particularly useful for researchers who need more than a basic summary of a scientific publication. Its workflow can search for relevant papers, retrieve evidence, assess the usefulness of individual passages, and generate an answer based on the selected information. It also supports agentic retrieval, allowing an AI agent to refine its searches and gather additional evidence when necessary. This makes it useful for scientific question answering, document analysis, literature synthesis, and the examination of conflicting research claims.

A key feature of PaperQA2 is its emphasis on evidence-based responses. Generated answers include citations that help users trace statements back to the documents used during the research process. Researchers can inspect the supporting material and decide whether the evidence is sufficient for their intended use. The tool also includes document metadata awareness, contextual summarisation, and ranking methods intended to improve the relevance of retrieved information. Its configurable architecture allows developers to adapt the system to different research collections and language-model providers.

PaperQA2 can also be used through a Python interface and command-line workflow. Developers can install the package, create searchable collections of research documents, and submit questions programmatically. It supports asynchronous operations and can be configured with different compatible language models and embedding providers. This makes it suitable for research teams that want to build their own scientific knowledge systems instead of relying exclusively on a hosted research interface.

FutureHouse has also developed Aviary, an open-source framework for training and evaluating language agents on challenging scientific tasks. The framework was created to make it easier to train AI agents using realistic research activities rather than relying only on general conversational benchmarks. Its applications cover tasks such as literature research, reasoning about DNA, and protein design. Aviary contributes to the broader goal of creating AI systems that can perform specialised scientific work and be evaluated against meaningful research objectives.

Another important project is ether0, a scientific reasoning model developed specifically for chemistry. FutureHouse introduced ether0 as a model trained with reinforcement learning to improve performance on chemistry-related reasoning and molecular design tasks. The project reflects the organisation’s interest in building models that understand specialised scientific problems rather than relying entirely on general-purpose language models. FutureHouse made ether0 available as an open-source research resource so that scientists and developers can examine and build on the work.

FutureHouse also conducts research into end-to-end AI-driven scientific discovery. One example is Robin, a multi-agent system developed to connect different stages of the research process. Robin was designed to generate hypotheses, plan experiments, analyse results, and use the findings to guide further scientific investigation. In one reported research project, the system identified ripasudil, an existing glaucoma drug, as a promising candidate for investigation in dry age-related macular degeneration. The work demonstrated how AI agents could contribute to a connected scientific workflow rather than simply answering isolated questions.

The organisation’s research also includes the development of scientific benchmarks. AI systems need reliable evaluation methods because strong performance on ordinary language tests does not necessarily mean that they can complete complex scientific tasks accurately. FutureHouse developed LAB-Bench to assess practical capabilities relevant to biological research. It also developed BixBench, which focuses on open-ended bioinformatics tasks. These benchmarks help researchers and AI developers evaluate how well models perform on scientific questions that require specialised knowledge, evidence retrieval, and analytical reasoning.

FutureHouse’s work extends beyond software and model development. The organisation operates an AI-for-Science Independent Postdoctoral Fellowship programme that pairs early-career researchers with advanced AI tools and academic co-advisors at established research institutions. Fellows lead research projects at the intersection of AI and science while receiving access to technical resources and scientific guidance. The programme is intended to develop researchers who can combine scientific expertise with modern AI methods and contribute to the future of computational research.

The organisation also develops learning resources and training initiatives to help academic scientists incorporate AI into their research workflows. These activities recognise that scientific AI tools are most useful when researchers understand how to apply them, assess their limitations, and interpret their outputs appropriately. Training and collaboration can help researchers develop practical skills in AI-assisted literature analysis, scientific reasoning, and computational research.

An important distinction is the relationship between FutureHouse and Edison Scientific. Edison Scientific is a for-profit company that spun out of FutureHouse in 2025 to commercialise AI tools for scientific research. The two organisations are independent and have different missions. FutureHouse continues to concentrate on nonprofit research and scientific discovery for the public good, while Edison Scientific develops and commercialises the technology through its own platform. FutureHouse’s official website explains that AI systems originally developed under the FutureHouse name are now part of Edison Scientific’s platform. Consequently, the original Crow, Falcon, Owl, Phoenix, and related platform capabilities should not automatically be treated as a currently operated standalone FutureHouse subscription service.

FutureHouse is supported through philanthropic partnerships and grants. Its early development received support from Eric and Wendy Schmidt, while additional support has come from organisations including Open Philanthropy, the National Science Foundation, and the AI Safety Institute. This funding model allows the organisation to pursue research initiatives and make selected scientific resources available to the broader community.

FutureHouse is best understood as an AI-for-science research organisation with open-source tools, scientific models, research programmes, and a history of developing specialised AI agents. Its work is relevant to biology, biomedical research, chemistry, computational science, and scientific software development. Researchers can explore its publications and open-source projects, while organisations interested in supporting specific research areas can engage with its partnership and fellowship initiatives.

Pricing

FutureHouse’s nonprofit research activities and selected open-source tools are available to the research community without a conventional software subscription. PaperQA2, Aviary, and ether0 are examples of projects made available for researchers and developers to use and build upon.

• PaperQA2 – Free: Open-source software available for installation and use. External language-model, embedding, and research-data services may have their own usage costs.
• Aviary – Free: Open-source framework for training and evaluating AI agents on scientific tasks.
• ether0 – Free: Open-source chemistry reasoning model made available for research and development.
• Research and publications – Free: Selected research papers, benchmarks, announcements, and learning resources can be accessed through the organisation’s website.
• Fellowship programme – Research programme: Selected early-career researchers receive access to AI tools, research resources, and academic co-advisors through the fellowship programme.
• Commercial AI research platform – Separate offering: Commercial access to technology developed by FutureHouse is handled through Edison Scientific, which operates independently.

Review

FutureHouse is an interesting organisation in the AI research space because it focuses on developing systems that can contribute to scientific discovery rather than simply generating general-purpose text. Its work addresses practical research challenges such as searching scientific literature, identifying previous experiments, analysing evidence, and supporting the development of new scientific hypotheses.

One of its strongest contributions is the development of specialised research agents. The original Crow, Falcon, Owl, and Phoenix agents were designed around different scientific tasks, making the platform more focused than a general-purpose chatbot. Their specialised roles offered researchers different ways to investigate scientific questions, conduct literature reviews, and explore chemistry-related problems.

PaperQA2 is another significant strength. Its ability to retrieve information from scientific documents and generate answers with in-text citations makes it useful for researchers who need traceable evidence. Its support for local document collections, configurable language models, and Python-based workflows also gives developers flexibility when building research applications.

FutureHouse’s open-source projects provide additional value. PaperQA2, Aviary, and ether0 give researchers and developers resources they can examine, adapt, and incorporate into their own work. The organisation’s scientific benchmarks, including LAB-Bench and BixBench, also contribute to the wider AI research community by supporting more specialised evaluation of scientific capabilities.

Its research into end-to-end scientific discovery is particularly promising. Projects such as Robin demonstrate an approach in which AI agents contribute to multiple stages of research, from generating hypotheses to analysing experimental findings. The postdoctoral fellowship programme is another positive feature because it connects AI development with working scientists and academic institutions.

However, FutureHouse is not a conventional all-in-one research application with a single current subscription interface. It is a nonprofit research laboratory, and its commercial AI platform technology has moved into Edison Scientific. This distinction may confuse users who discover older references to the FutureHouse Platform and expect to access the original agents through the same service.

Another limitation is that using its open-source tools can require technical knowledge. PaperQA2, for example, involves software installation, configuration, document indexing, and access to compatible language-model services. Researchers may also need to manage external API credentials and associated usage costs. These requirements can make the tools less convenient for people who prefer a simple browser-based research application.

AI-generated scientific answers also require careful evaluation. Citation grounding and evidence retrieval can make research outputs easier to inspect, but they do not guarantee that every interpretation or conclusion is correct. Researchers should verify important claims against the original publications and use appropriate scientific judgement before relying on generated material in research decisions.

FutureHouse’s work is also specialised. Its strongest focus is biology and complex scientific research, so it is not intended to serve as a general writing assistant, business chatbot, or everyday productivity application. Users outside scientific research may find more suitable tools designed for their specific needs.

FutureHouse is a valuable organisation for the scientific AI community, particularly through its open-source research tools, specialised agent development, benchmarks, and fellowship programme. It is most relevant to researchers and developers interested in AI-assisted scientific discovery. Anyone evaluating it as a software product should distinguish FutureHouse’s nonprofit research activities from the commercial services now operated by Edison Scientific.