STORM is a free research prototype from Stanford's OVAL lab that researches a topic on the web and writes a long, Wikipedia-style report with citations. It runs as a hosted demo at storm.genie.stanford.edu, with a sign-in and a Co-STORM mode where you can join the AI's research conversation, and as open-source code under the MIT license with about 31,500 stars on GitHub. There is no paid plan. The code's last commit was on September 30, 2025 and its last release in January 2025, so treat it as an academic tool rather than a maintained product.
This analysis is based on documentation, public user reports, and vendor materials — not yet on our own hands-on testing. How we rate
STORM comes from Stanford University's Open Virtual Assistant Lab, and its name spells out the method: Synthesis of Topic Outlines through Retrieval and Multi-perspective Question Asking. Give it a topic and it researches the web, builds an outline and writes a long report with citations, in the style of a Wikipedia article. The researchers presented it at NAACL 2024, followed by a collaborative version, Co-STORM, at EMNLP 2024.
Two things set expectations. The website at storm.genie.stanford.edu calls STORM "a research prototype that supports interactive knowledge curation", and its terms describe it as a research preview with limited safety measures that may generate offensive content. It is free, runs on cloud credits donated by Microsoft Azure, and requires a sign-in. There is no company behind it, no paid plan and no support desk.
For what it does, that is fine. STORM is one of the more interesting free research tools, and it arrived months before the deep research features that commercial assistants now ship.
Most chatbots answer a question from a single point of view. STORM starts by asking what different people would want to know about the topic. It simulates conversations in which writers with different perspectives, say a historian, an engineer and an economist, question a topic expert whose answers are grounded in search results. Those conversations surface angles a single prompt would miss.
From the gathered material, STORM drafts an outline, then writes the article section by section with citations to the sources it retrieved. The result reads like a first draft of an encyclopedia entry: broad, organized and referenced. It is well suited to getting oriented on an unfamiliar subject, planning a longer piece of writing or finding the sources worth reading first.
The limits come with the format. The reports aim for breadth rather than argument, they reflect what the search step found, and the text can misstate a source even when the citation is real. Treat a STORM report as a map of a topic, not a finished piece of work.
Co-STORM, now the headline on the website, turns the process into something you can watch and steer. AI agents representing different perspectives discuss your topic while you observe, and you can step in with questions to push the research in a direction you care about. When you have seen enough, you ask for the report.
The design goal, described in the lab's paper, is learning about "unknown unknowns": the aspects of a subject you would not think to ask about. For students and researchers approaching a new field, that makes Co-STORM more useful than the one-shot report, since the conversation itself teaches the shape of the topic.
The code lives in the stanford-oval/storm repository under the MIT license, with around 31,500 stars and 3,000 forks, and installs as the knowledge-storm Python package. Running it yourself means supplying the language model and search provider through your own API keys, so you control which models it uses, which sources it searches and how the output is stored.
That flexibility suits developers and research groups who want to adapt the method, for example to run it over internal documents or with a specific model. The costs are yours: every report makes many model and search calls, and the total depends on the models chosen.
The project's activity has slowed sharply. As of September 30, 2026, the repository's last commit was exactly a year old, the last tagged release was version 1.1.0 from January 2025, and around 111 issues were open. The hosted demo still works and the website carries a 2026 copyright, but new models, integrations and fixes are not arriving the way they do in commercial tools.
That does not make STORM useless; the method still works and the code is open. It does mean nobody promises it will keep running, and anyone building on the code should expect to maintain their own fork.
For a free, structured overview of a topic with sources, STORM is worth a try, and Co-STORM is a genuinely different way to explore a subject. For ongoing research work, maintained tools do more. Gemini includes Deep Research even on its free plan, ChatGPT and Perplexity offer deep research with current models and follow-up questions, and NotebookLM works from sources you choose.
Academic literature needs specialized tools: Elicit and Consensus search peer-reviewed papers, and Scite shows how papers have been cited. STORM's niche is the reader who wants a Wikipedia-style starting point on anything, for free, and does not mind using an academic prototype to get it.
No. The hosted research preview at storm.genie.stanford.edu is free with a sign-in, and the code is open source under the MIT license. Stanford thanks Microsoft Azure for the cloud credits that run the demo. If you run STORM yourself, you pay for the language model and search APIs it calls.
A long, structured report on your topic in the style of a Wikipedia article: an outline, several sections and citations to the web sources it used. It gets there by simulating conversations between writers with different perspectives and a topic expert grounded in search results, then turning what they found into an outline and an article.
Co-STORM is the collaborative mode. Instead of waiting for a finished report, you watch AI agents with different perspectives discuss the topic, step in with your own questions to steer the research, and request the report when you are ready. It is designed for learning about a subject, including the questions you did not know to ask.
Barely. The hosted demo was online in September 2026, but the GitHub repository's last commit was on September 30, 2025, the last tagged release was version 1.1.0 in January 2025, and about 111 issues were open. It remains a popular research project with around 31,500 stars, but it is not actively developed like a commercial tool.
Yes, if you are comfortable with Python. The project is published as the knowledge-storm package, and you configure the language models and a search provider with your own API keys. Costs depend on the models and how many reports you generate, and you get full control over the setup, sources and output.
Deep research features in ChatGPT, Gemini and Perplexity are maintained products with current models, file support and conversation afterwards, and Gemini includes Deep Research even on its free plan. STORM is a research prototype focused on one thing, a Wikipedia-style report built through multi-perspective questioning, and its Co-STORM mode lets you watch and steer that process.
They point to real sources it retrieved, which makes checking easier, but the text can still misstate what a source says. The demo's terms describe STORM as a research preview with limited safety measures. Use its reports as a map of a topic and verify important claims in the cited sources before relying on them.
Cited AI search plus the new Computer agent: free plan, Pro at $20 or $200 a year, Max at $200, Computer credits, and the free Comet browser.
An academic search engine that answers from peer-reviewed papers. Basic search is free, Pro starts at $12 a month billed yearly, and students get up to 40% off.
A free chat app and one of the cheapest frontier-class APIs: V4.1-Flash with vision and open weights, from $0.15 per million input tokens off-peak.
