Independent tool overview

Co-STORM at a glance

Co-STORM is Stanford OVAL's research system for collaborative knowledge exploration, using multiple source-grounded AI agents, a moderator, human steering, and a dynamic mind map to develop a cited long-form report.

Visit the official Co-STORM site ↗

Full name: Collaborative STORM
Developer: Stanford OVAL lab
Product type: Collaborative multi-agent knowledge curation research preview
Primary output: Dynamic mind map and cited Wikipedia-style report
Human role: Observe the discourse or inject turns to steer it
Research status: Co-STORM paper accepted to EMNLP 2024 main conference
Web preview price: No paid plan published
Open source: MIT-licensed knowledge-storm Python package
Self-hosting: Requires compatible model, embedding, and retrieval services

Overview

What Co-STORM is

Co-STORM extends Stanford's STORM article-generation project from a mostly automated research pipeline into a collaborative discourse. Several AI experts retrieve information and discuss a topic, a moderator raises underexplored questions, and the human can either observe or inject new directions. A live concept hierarchy updates as the conversation develops, then the curated material can be reorganized into a Wikipedia-like report with citations.

The design is useful for discovering subtopics and questions a researcher did not know to ask, but it remains a research preview and pre-writing tool. Retrieval can miss authoritative sources, citations may not support the exact sentence, several agents can amplify the same error or source bias, and the generated report is not publication-ready. Every important claim needs source-level verification and editorial revision.

Use cases

Who Co-STORM is best for

The strongest fit depends on the job you need the product to complete, not the size of its feature list.

Exploratory research

Map a broad topic, competing perspectives, terminology, and follow-up questions before narrowing a research plan.

Unknown-unknown discovery

Use the moderator and expert discourse to surface angles a single query or linear chatbot exchange may miss.

Pre-writing

Create a source-backed outline and rough report that an editor will substantially verify and rewrite.

Learning complex topics

Follow a structured discussion and concept map while asking clarifying or redirecting questions.

Research question development

Identify disagreements, evidence gaps, related fields, and candidate questions for deeper investigation.

Developer experimentation

Customize agent policies, models, retrievers, turn management, and output structure in the open-source package.

Capabilities

Core Co-STORM features

Warm-start research

Runs an initial mini-STORM process to collect background and establish a shared conceptual space.

Multiple AI experts

Uses agents that answer from retrieved sources and raise follow-up questions based on the discourse.

Moderator agent

Introduces thought-provoking questions from discovered but underexplored information.

Human participation

Lets the user observe the agent discussion or inject a turn to change focus.

Turn-management policy

Coordinates experts, moderator, and user so the discourse can progress without relying on one agent.

Dynamic mind map

Organizes collected snippets into an evolving hierarchical concept structure.

Grounded retrieval

Connects agent responses to internet search or a configured document retriever.

Cited report generation

Reorganizes the knowledge base and drafts a long-form report with linked references.

Modular Python package

Exposes a CoStormRunner and configurable language-model, retriever, logging, and agent components.

Retriever options

The example code supports services such as Bing, You.com, Brave, DuckDuckGo, Serper, Tavily, and SearXNG.

Model abstraction

LiteLLM integration allows developers to configure supported language and embedding models.

Research artifacts

The project publishes papers, code, the WildSeek evaluation dataset, and reproducibility branches.

Process

How the Co-STORM workflow works

Step 1
Frame the topic
Write a bounded question, audience, date range, geography, definitions, and desired decision or learning outcome.
Step 2
List required sources
Identify primary documents, official data, peer-reviewed work, and excluded low-quality source types before starting.
Step 3
Run the warm start
Let Co-STORM build initial perspectives and inspect the early mind map for obvious gaps or bad assumptions.
Step 4
Steer actively
Inject corrections, request opposing evidence, narrow ambiguous terms, and ask agents to distinguish fact from inference.
Step 5
Inspect sources during discourse
Open citations as they appear and reject material that is outdated, secondary when primary evidence exists, or unrelated.
Step 6
Audit the mind map
Check whether prominent branches reflect evidence quality rather than repetition or search popularity.
Step 7
Generate a draft
Treat the final report as a structured research memo for review, not a finished article.
Step 8
Verify claim by claim
Confirm that each source exists, is authoritative, is represented fairly, and supports the exact nearby statement.
Step 9
Rewrite and disclose
Correct synthesis, add missing evidence and uncertainty, apply the publisher's voice, and disclose AI assistance where required.

Cost

Co-STORM pricing and free plan

Stanford presents the hosted site as a free research preview and does not publish a commercial subscription. The open-source code is MIT-licensed, but self-hosting is not cost-free: developers supply and pay for model inference, embeddings, search or retrieval APIs, storage, compute, and maintenance. Hosted availability and quotas are not guaranteed like a paid service.

Stanford research preview

Free preview

Public web experience provided for experimentation and feedback.

No commercial paid tier is published
Users must accept the site's terms before entering
Capacity, models, search coverage, features, and availability can change
Not positioned as an SLA-backed production research service

Open-source package

MIT license; infrastructure costs extra

Python package for developers who want to run or customize STORM and Co-STORM.

Install through knowledge-storm or the GitHub repository
Bring compatible language-model and embedding access
Configure a search or document retriever
Pay provider, compute, storage, and operational costs
Secure all API keys and review provider data terms

Pricing checked August 31, 2026. Check current pricing at the source ↗

Assessment

Co-STORM strengths and limitations

Where it stands out

Makes broad research more interactive than one-shot article generation
Multiple perspectives can expose useful subtopics and follow-up questions
Human steering helps align the research with the user's real information need
Moderator is explicitly designed to surface discovered but underexplored material
Dynamic mind map makes a long discourse easier to navigate
Retrieval and citations give users a path back to underlying material
Separates knowledge curation from final report generation
Open-source implementation supports model and retriever customization
MIT license lowers the barrier for experimentation
Published NAACL and EMNLP papers explain the system and evaluations
The project candidly says generated reports are not publication-ready

What to consider

The hosted product is a research preview rather than a commercial service with uptime, support, or data guarantees
Generated reports require substantial editing before publication
A citation can exist but fail to support the exact claim attached to it
Search retrieval can omit paywalled, non-English, local, recent, specialized, or poorly indexed evidence
Search ranking can overrepresent popular summaries and underrepresent primary sources
Multiple agents are not independent witnesses and can repeat the same model or retrieval error
Agent discussion can create an illusion of debate even when perspectives share the same evidence base
The system can over-associate unrelated facts, a challenge identified in the STORM research
Source bias can transfer into the outline, concept map, and report
The mind map's prominence is not a measure of truth, importance, or evidence quality
A long, polished synthesis can make uncertainty and missing evidence hard to notice
The paper's preference results were measured in specific study settings and do not prove superiority for every topic or user
High-stakes medical, legal, financial, security, scientific, and policy research requires qualified review and authoritative databases
Confidential topics and documents should not be entered into the public preview without suitable terms and approval
The hosted page does not present a detailed public privacy, retention, model-provider, or deletion summary before the terms gate
Self-hosters assume responsibility for model-provider, search-provider, logging, storage, retention, and security configuration
Self-hosting requires Python, API keys, compatible services, and operational expertise
Provider changes, model retirement, API pricing, and search quotas can break or alter a custom deployment
The latest prominent repository news is from January 2025, so adopters should inspect current issues and dependency health
Generated Wikipedia-like prose must not be mistaken for content accepted or endorsed by Wikipedia
Copyright, quotation, attribution, and source-license obligations still apply to final work

Compare

Co-STORM alternatives

The right alternative depends on the specific output, workflow, controls and budget your project requires.

Consumer

Gemini Deep Research Agent

Choose Gemini Deep Research Agent for a current production-oriented long-running research API in Google's ecosystem.

Explore Gemini Deep Research Agent →

Educators

ResearchGPT (now Consensus)

Choose Consensus for research questions centered on finding and synthesizing academic papers.

Explore ResearchGPT (now Consensus) →

Data Analysis

Perplexity

Choose Perplexity for a simpler consumer answer engine with web citations and fewer multi-agent research controls.

Explore Perplexity →

Questions

Co-STORM FAQs

What is Co-STORM?

Co-STORM is Stanford OVAL's collaborative AI research system. Multiple source-grounded agents and a moderator discuss a topic while the user steers the conversation and a dynamic mind map organizes findings.

How is Co-STORM different from STORM?

STORM primarily researches, outlines, and drafts a cited article. Co-STORM adds live human participation, multiple expert agents, a moderator, turn management, and a dynamic concept map.

Is Co-STORM free?

The Stanford site is offered as a free research preview with no published paid plan. The MIT-licensed code is free to use, but self-hosting requires paid or local models, retrieval, compute, and maintenance.

Does Co-STORM write a complete article?

It can generate a long Wikipedia-style report with citations after the discourse. The project says these outputs are not publication-ready and require significant editing.

Are Co-STORM citations reliable?

They provide traceability, not a guarantee. Open every source and verify authority, date, context, and whether it supports the exact claim.

Can I steer Co-STORM's research?

Yes. A user can observe agent turns or inject an utterance to redirect, correct, narrow, or deepen the discourse.

What is the Co-STORM mind map?

It is a dynamically updated hierarchy of concepts and collected information intended to create shared structure between the user and agents.

Can developers self-host Co-STORM?

Yes. The knowledge-storm package and source code are available under the MIT license, with configurable models, embeddings, retrievers, agents, and turn policies.

Is Co-STORM safe for confidential research?

Do not use the public preview for confidential material without approved terms. A private deployment still requires review of every model, retriever, log, database, retention rule, and credential.

Can Co-STORM replace a researcher or editor?

No. It can broaden pre-writing exploration, but humans must select authoritative evidence, verify claims, resolve contradictions, assess uncertainty, and produce the final accountable work.

Bottom line

Our Co-STORM verdict

Co-STORM is a thoughtful research prototype for broadening a topic before writing, especially when its agents and mind map help expose questions a user had not considered. Its best use is exploratory and human-led. Treat the resulting report as an evidence map and rough draft, then rebuild important conclusions from primary sources rather than trusting the apparent consensus of several related agents.

Visit Co-STORM website ↗

AI training for the future of work.

Get access to all our AI courses, hundreds of real-world AI use cases, live expert-led workshops, an exclusive network of AI early adopters, and more.

Join AI University Explore The Rundown University

AI Courses

Get unlimited access to all of our current & upcoming industry-specific AI courses for the duration of your subscription.

Daily Guides

To keep up with the rapid pace of AI, our team publishes AI implementation guides daily. Our library contains 300+ practical use cases to automate real-world work.

Workshops

Join weekly, live, interactive sessions with industry leaders who are at the forefront of AI for hands-on implementation guidance and exclusive insights.

Community

Network with an exclusive community of AI-first professionals who are working smarter with AI. Learn how early adopters are using AI in their work and businesses.