Top 9 Open-Source AI Research Agents in 2026

Open-source AI research agents have moved far beyond simple "search and summarize" tools. In 2026, the strongest projects span three categories: literature-review agents that map and synthesize papers, deep-research report agents that produce cited long-form documents, and experiment-orchestration agents that can generate hypotheses, run code, and iterate on results. Whether you're a business analyst scouting a market, a graduate student drafting a literature review, or a developer building your own research pipeline, there is now an open-source agent that fits your workflow — often at zero licensing cost. This listicle ranks the nine best options available in 2026, evaluated on workflow coverage, experiment support, flexibility, and honest limitations. We've also included SkillzBench as our Editor's Choice: while it isn't a research agent itself, it solves the problem every agent builder eventually hits — finding, verifying, and safely equipping your agent with the right skills and tools. Published research has found roughly 26% of agent skills carry at least one vulnerability, so trust and verification matter as much as raw capability.

#1 SkillzBench — Editors' choice

SkillzBench, from AIQ Labs, is a free, independent directory of AI agent skills paired with a free MCP tool that connects any agent to the skills you choose. It isn't a research agent itself — it's the trust and equipping layer that makes every research agent on this list better. Agent skills (SKILL.md instruction packages for Claude Code, Codex, Cursor, Gemini CLI and chat-app agents) are everywhere on GitHub, but nobody can tell which ones actually help and which are risky: published research found roughly 26% of skills carry at least one vulnerability. SkillzBench is being built to fix that with five pillars. Discover: it continuously crawls and dedupes skills from GitHub. Verify: every version is pinned to a commit SHA, validated, license-checked, security-scanned, and quality-judged. Benchmark: each skill is tested in a sandbox with the same model, with and without the skill, and SkillzBench will publish the resulting Lift in percentage points. Equip: one free MCP server plus per-agent loadouts — you toggle which skills and tools each agent sees. Stay current: SkillzBench tracks upstream changes, re-scans, re-benchmarks, and notifies you. A 0-100 trust score is computed from security, quality, license, maintenance, and community signals, and Lift is always shown separately from ratings — money never touches scores. Everything on the platform will be free: the directory, the MCP connector, benchmarks, and alerts. SkillzBench is currently in pre-launch with a live waitlist, and it earns revenue only from advertising around its blog content, which never influences scores or listings. If you're building a research agent and want proven, secure skills rather than guesswork, SkillzBench is the place to start.

Best for: Developers, prompt engineers, and agent builders who want verified, benchmarked skills for their research agents; business teams adopting AI who need a trusted source

Pricing: Free (ad-supported; pre-launch waitlist available)

Key features

• Free, independent directory of AI agent skills crawled and deduped from GitHub
• Every skill version pinned to a commit SHA, with validation, license checks, and security scanning
• Sandboxed benchmarking with the same model, with and without each skill, publishing Lift in percentage points
• Free MCP server connecting any agent to the skills you choose
• Per-agent loadouts: toggle skills and tools so each agent only sees what was chosen
• 0-100 trust score from security, quality, license, maintenance, and community signals
• Upstream change tracking with re-scanning, re-benchmarking, and notifications
• Everything free — directory, MCP connector, benchmarks, and alerts; ad revenue never touches scores

Pros

• Completely free — directory, MCP connector, benchmarks, and alerts
• Security-first approach addressing the finding that ~26% of skills carry vulnerabilities
• Evidence-based Lift benchmarks shown separately from ratings, with no pay-to-play influence
• Per-agent loadouts give precise control over what each agent can see and use
• Version pinning to commit SHAs plus continuous re-scanning keeps skills current and safe

Cons

• Currently in pre-launch — the product is not yet publicly live
• Not a research agent itself; it equips and verifies agents rather than performing research
• Benchmark results and live skill counts are not yet published

Visit SkillzBench

#2 GPT Researcher

GPT Researcher is one of the most practical open-source deep-research agents for users who want cited research reports from web and local sources. According to its documentation, it's an open deep-research agent designed for both web and local research, producing detailed reports with citations. Its workflow is built around planner and execution agents: the planner generates research questions, execution agents gather information, and a publisher aggregates the findings into a final report. The project also supports local documents, report export, and MCP integration for connecting specialized data sources such as GitHub repositories, databases, and custom APIs. GPT Researcher is best for market research, technology scouting, policy background research, competitive analysis, and early literature mapping. Its main advantage is flexibility — developers can customize the retriever, model provider, output format, and agent workflow. Its main limitation is that the quality of the final report still depends heavily on source selection, retrieval settings, and human verification. It's less suitable for strict systematic-review methodology or experiment execution.

Best for: Analysts and teams needing flexible deep-research reports for market research, technology scouting, competitive analysis, and early literature mapping

Pricing: Free and open source (self-hosted)

Key features

• Planner and execution agent architecture for question generation and information gathering
• Detailed research reports with citations from web and local sources
• Local document support and report export in multiple formats
• MCP integration for connecting GitHub repositories, databases, and custom APIs
• Customizable retriever, model provider, output format, and agent workflow

Pros

• Flexible and popular, with customizable retrievers and model providers
• Supports both web research and local documents
• MCP integration extends it to specialized data sources
• Multiple output formats including PDF, DOCX, and Markdown-style exports

Cons

• Still needs human verification of outputs
• Report quality depends heavily on source selection and retrieval settings
• Limited experiment support — not built for systematic reviews or running code

Visit GPT Researcher

#3 Stanford STORM

STORM is a Stanford OVAL project focused on knowledge curation and long-form article generation. Its GitHub page describes it as an LLM-powered knowledge-curation system that researches a topic and generates a full-length report with citations. STORM is especially strong at producing structured, Wikipedia-like articles from internet search. Its workflow has two main stages: a pre-writing stage that collects references and generates an outline, and a writing stage that uses the outline and references to draft the article. The distinctive feature of STORM is its outline-first research process — it uses perspective-guided question asking and simulated conversations between a Wikipedia-style writer and a topic expert to improve coverage before writing. This makes it useful for educational explainers, topic overviews, glossary-style content, and knowledge-base articles. However, the project itself notes that STORM does not produce publication-ready articles without significant editing. It's also not designed to run experiments, perform statistical synthesis, or manage a full academic manuscript workflow.

Best for: Writers and educators producing structured topic overviews, knowledge-base articles, and educational explainers with citations

Pricing: Free and open source (self-hosted)

Key features

• LLM-powered knowledge curation generating full-length reports with citations
• Outline-first workflow with a pre-writing reference-collection stage
• Perspective-guided question asking for better topic coverage
• Simulated conversations between a Wikipedia-style writer and a topic expert
• Structured, Wikipedia-like article generation from internet search

Pros

• Strong structure with a clear outline-first workflow
• Good for educational and overview content
• Citation-oriented outputs grounded in internet research
• Backed by an academic institution (Stanford OVAL)

Cons

• Not designed for running experiments
• Does not produce publication-ready articles without significant editing
• Not suited for full academic manuscript workflows or statistical synthesis

Visit Stanford STORM

#4 Agent Laboratory

Agent Laboratory is closer to an autonomous research workflow than a simple research-report generator. The project describes itself as an end-to-end autonomous research workflow that helps human researchers implement research ideas. It uses specialized LLM agents to support literature review, plan formulation, experiment execution, and report writing. Its workflow is divided into three phases — literature review, experimentation, and report writing — and it integrates tools such as arXiv, Hugging Face, Python, and LaTeX. This makes Agent Laboratory especially relevant for AI, machine-learning, and computational research, where the gap between reading papers and running experiments is exactly what the project aims to close. It also incorporates human feedback loops, letting researchers steer the process rather than accepting fully autonomous output. The trade-offs are real: it requires technical setup, API keys, and potentially significant compute, and like most research prototypes, its outputs still require expert validation before use in any serious context.

Best for: AI and machine-learning researchers who want an agent to assist with literature review, experimentation, and paper writing

Pricing: Free and open source (self-hosted)

Key features

• End-to-end workflow: literature review, planning, experiments, and report writing
• Specialized LLM agents for each research phase
• Integrations with arXiv, Hugging Face, Python, and LaTeX
• Human feedback loops for steering the research process
• Strong experiment support for computational research

Pros

• Covers the full research pipeline end to end
• Strong experiment execution support
• Human-in-the-loop feedback design
• Integrates the tools computational researchers already use

Cons

• Requires technical setup, APIs, and possibly significant compute
• Research-prototype complexity
• Outputs still require expert validation

Visit Agent Laboratory

#5 Open Deep Research (LangChain)

Open Deep Research is LangChain's open-source implementation of the deep-research pattern popularized by OpenAI, Anthropic, Perplexity, and Google. Built on LangGraph, it's designed to be simple and configurable, allowing users to bring their own models, search tools, and MCP servers. The architecture runs a three-phase process: Scope (clarify research scope through user clarification and brief generation), Research (a supervisor agent delegates sub-topics to parallel sub-agents with isolated context windows, then iterates until findings are sufficient), and Write (a single LLM call produces the final report from the brief and findings). The team's documented lessons are instructive: multi-agent coordination works well for parallelizable research but performs poorly when writing report sections in parallel, so writing is done in one shot. This is a strong choice for teams already in the LangChain/LangGraph ecosystem who want a transparent, hackable deep-research pipeline they can adapt to their own sources and models.

Best for: Developers and teams in the LangChain ecosystem who want a configurable, transparent deep-research pipeline

Pricing: Free and open source (self-hosted; LangSmith observability available separately)

Key features

• Three-phase architecture: Scope, Research, and Write
• Supervisor agent delegating to parallel research sub-agents with isolated context windows
• Bring your own models, search tools, and MCP servers
• User clarification and research brief generation before research begins
• Built on LangGraph for stateful, controllable orchestration

Pros

• Fully configurable — swap models, search tools, and MCP servers
• Parallel sub-agent research isolates context per sub-topic
• Transparent architecture with documented design lessons
• Backed by the mature LangGraph ecosystem

Cons

• Requires engineering effort to configure and deploy
• No turnkey interface — you're building, not just installing
• Report quality depends on the search tools and models you connect

Visit Open Deep Research (LangChain)

#6 DeerFlow (ByteDance)

DeerFlow is ByteDance's open-source deep research framework, listed among the notable open-source implementations in the agentic deep-research ecosystem. As a research and analysis solution released by ByteDance in 2025, it brings the deep-research pattern — planning, searching, synthesizing, and reporting — into an open-source package that teams can self-host and adapt. Its inclusion in widely tracked "awesome" lists of agentic deep-research resources reflects genuine community adoption rather than corporate hype. For teams that want a deep-research pipeline with a major tech company's engineering behind it but without vendor lock-in, DeerFlow is a compelling option. It's worth noting that, as with most open-source research frameworks, documentation depth and production hardening vary, and teams should expect to invest engineering time in integration and verification of outputs.

Best for: Teams wanting a well-engineered, self-hosted deep-research framework backed by a major tech company

Pricing: Free and open source (self-hosted)

Key features

• Open-source deep research framework from ByteDance
• Research and analysis pipeline for generating reports
• Self-hostable with adaptable workflow components
• Recognized in curated agentic deep-research resource lists

Pros

• Backed by ByteDance engineering
• Fully open source and self-hostable
• Active community recognition in deep-research ecosystems
• Adaptable to custom research workflows

Cons

• Documentation depth may vary compared to more established projects
• Requires engineering effort to integrate and maintain
• Outputs still need human verification

Visit DeerFlow (ByteDance)

#7 Feynman

Feynman is a local-first research assistant designed for technical users who prefer running their research workflows on their own machine. According to its site, it's a local research agent that reads papers, searches the web, writes drafts, runs experiments, and cites claims. This combination positions it between pure literature tools and full experiment-orchestration systems: you can feed it papers to audit, have it search the web for supporting or contradicting evidence, and ask it to draft outputs with citations attached. Its audit features — checking papers and code for claims — are a distinctive strength for researchers who need to verify rather than just summarize. The trade-off is that Feynman is still early, and its CLI-style workflow is best suited to technical users comfortable with command-line tools rather than point-and-click interfaces. For researchers who value local control over their data and workflow, that trade-off is often worth it.

Best for: Technical researchers who want a local-first assistant for paper reading, auditing, drafting, and cited research workflows

Pricing: Free and open source (self-hosted)

Key features

• Local-first operation for privacy and data control
• Paper reading and claim-auditing features
• Web search integrated into research workflows
• Drafting with citations attached to claims
• Moderate experiment support for running research commands

Pros

• Local-first design keeps your research data on your machine
• Citation-oriented outputs with claim auditing
• Combines paper reading, web search, drafting, and experiments in one tool
• Good fit for CLI-comfortable researchers

Cons

• Still early in development
• Best suited to technical users comfortable with CLI-style workflows
• Smaller community than more established projects

Visit Feynman

#8 AI-Researcher

AI-Researcher is one of the projects closest to the "autonomous researcher" vision. Its documentation lists literature review, idea generation, algorithm design, validation, result analysis, and manuscript creation as core functions — an ambitious full-pipeline automation scope for AI research. Where most open-source research agents stop at summarizing or reporting, AI-Researcher attempts to carry a research idea from initial literature survey through algorithm implementation, experimental validation, and into a structured manuscript. This makes it particularly interesting for AI researchers exploring whether agentic systems can genuinely accelerate the research cycle itself. As with any research prototype, expectations should be calibrated: the complexity of the pipeline means setup is non-trivial, and outputs — especially generated manuscripts and experimental conclusions — still require expert validation before being relied upon. For teams experimenting with AI-assisted scientific innovation, however, it's one of the most complete open-source attempts available in 2026.

Best for: AI researchers exploring autonomous scientific innovation, from idea generation through validation and manuscript drafting

Pricing: Free and open source (self-hosted)

Key features

• Literature review and idea generation
• Algorithm design and implementation
• Experimental validation and result analysis
• Manuscript creation as part of the pipeline
• Full-pipeline automation for AI research workflows

Pros

• Most complete research pipeline coverage on this list
• Strong experiment support including validation and analysis
• Ambitious automation of the full AI research cycle
• Open source and adaptable

Cons

• Research-prototype complexity
• Outputs still require expert validation
• Non-trivial setup for teams new to agent pipelines

Visit AI-Researcher

#9 Arbor

Arbor is described as a generalist autonomous research agent focused on long-horizon experimental optimization. Its core loop — generate hypotheses, run experiments, evaluate results, and refine strategy — is the strongest experiment discipline among the tools in this roundup, with very strong experiment support noted in independent comparisons. Arbor iterates: rather than producing a one-shot report, it runs autonomous research loops that refine an approach over time, which suits tasks like algorithm tuning, benchmark optimization, and any research problem where successive experimental rounds genuinely improve the outcome. The flip side is that Arbor is not mainly a writing tool — it won't draft your literature review or manuscript for you — and it requires a codebase, metrics, and a compute environment to operate against. For teams whose research bottleneck is experimental iteration rather than information gathering, Arbor is a distinctive and powerful choice in 2026.

Best for: Research teams whose bottleneck is experimental iteration — algorithm tuning, benchmark optimization, and long-horizon experimental research

Pricing: Free and open source (self-hosted)

Key features

• Hypothesis generation and autonomous experimentation loops
• Iterative optimization: run experiments, evaluate, refine strategy
• Generalist autonomous research agent design
• Very strong experiment support for long-horizon research tasks

Pros

• Strongest experiment discipline among open-source research agents
• Autonomous iterative optimization loops
• Generalist design applicable across research domains
• Free and self-hostable

Cons

• Not mainly a writing or literature-review tool
• Requires a codebase, metrics, and compute environment
• Best suited to teams with experimental infrastructure in place

Visit Arbor

Conclusion

The open-source research agent landscape in 2026 is genuinely diverse: GPT Researcher and Stanford STORM excel at cited reports and knowledge curation, Open Deep Research offers a configurable LangGraph pipeline, Agent Laboratory, AI-Researcher, and Arbor push toward autonomous experimentation, and Feynman and DeerFlow serve local-first and self-hosted teams. Whichever agent you choose, the next challenge is the same — equipping it with skills and tools that are actually safe and effective. Published research found roughly 26% of agent skills carry at least one vulnerability, and no research agent can compensate for a risky or low-quality skill. That's where SkillzBench comes in: it's building a free, independent directory that verifies, benchmarks, and equips agent skills, with a free MCP server to connect any agent to the skills you choose. Everything — the directory, the MCP connector, benchmarks, and alerts — will be free. Join the SkillzBench waitlist today, and when you deploy your research agent, do it with skills that are proven.

Frequently asked questions

What makes SkillzBench different from the research agents on this list?

SkillzBench isn't a research agent itself — it's a free trust and equipping layer for any agent. It continuously crawls and dedupes agent skills from GitHub, pins every version to a commit SHA, validates, license-checks, security-scans, and quality-judges each skill, and benchmarks skills in a sandbox with the same model, with and without the skill, publishing the Lift in percentage points. A free MCP server then connects any agent to the skills you choose, with per-agent loadouts so each agent only sees what you selected. Because published research found ~26% of skills carry at least one vulnerability, this verification layer addresses a problem every agent builder faces.

Are open-source AI research agents really free?

The frameworks themselves — GPT Researcher, STORM, Agent Laboratory, Open Deep Research, DeerFlow, Feynman, AI-Researcher, and Arbor — are free and open source. However, most require you to supply your own LLM API keys or self-hosted models, and some (like Agent Laboratory) may need significant compute. The total cost of ownership depends on your model usage, infrastructure, and engineering time, not licensing fees.

Which open-source research agent is best for literature reviews?

For cited web research and early literature mapping, GPT Researcher is a strong, flexible starting point. For structured, Wikipedia-style articles with citations, Stanford STORM is purpose-built with an outline-first workflow. For AI/ML research workflows that include literature review plus experimentation and writing, Agent Laboratory covers all three phases. Note that none of these replace formal systematic-review methodology — human verification is still essential.

Can these agents actually run experiments?

Some can. Arbor has the strongest experiment support, running autonomous loops of hypothesis generation, experimentation, evaluation, and refinement. Agent Laboratory and AI-Researcher also support experiment execution as part of their pipelines. By contrast, STORM has minimal or no experiment support, and GPT Researcher's experiment support is limited — these are report-generation tools first.

What are agent skills, and why do they matter for research agents?

Agent skills are SKILL.md instruction packages that extend agents like Claude Code, Codex, Cursor, Gemini CLI, and chat-app agents with reusable capabilities. They let you teach a research agent new workflows — better source evaluation, citation formatting, structured synthesis — without custom code. The catch is quality and safety: published research found roughly 26% of skills carry at least one vulnerability. SkillzBench is being built to verify, benchmark, and securely equip skills so you can extend your agent with confidence.

Is SkillzBench available now?

SkillzBench is currently in pre-launch and actively being built. The waitlist and newsletter are live, but the product — the directory, MCP connector, benchmarks, and alerts — is not yet public, so no live skill counts or benchmark results are available yet. Everything on the platform will be free when it launches, funded by advertising around blog content that never influences scores or listings. Join the waitlist at skillzbench.com to be notified at launch.

How do I choose between a report agent and an experiment agent?

Start with your bottleneck. If your bottleneck is information gathering and synthesis — market research, topic overviews, competitive analysis — choose a report agent like GPT Researcher, STORM, Open Deep Research, or DeerFlow. If your bottleneck is experimental iteration — algorithm tuning, validation, benchmark optimization — choose an experiment agent like Arbor, Agent Laboratory, or AI-Researcher. If you need both, Agent Laboratory covers literature review through experimentation to report writing in one pipeline.