A curated list of open-source projects that automate scientific research — from literature review to idea generation, experiment execution, paper writing, and peer review.

 ·  English  ·  中文  ·  Browse tools  ·  Contribute

_{📅 Star counts last verified: 2026-10-13}

Multi-stage systems that autonomously handle several parts of the research loop, such as hypothesis generation, experimentation, analysis, and manuscript preparation.

Project & Stars	Description	Framework & LLM APIs
autoresearch	By Andrej Karpathy. A minimal single-GPU research harness where an external coding agent repeatedly edits train.py and runs fixed five-minute nanochat experiments under instructions in program.md.	Tools: Custom (PyTorch, nanochat) LLM APIs: External coding agents such as Anthropic Claude Code and OpenAI Codex
ARIS	Research workflow skills with independent cross-model review loops, persistent research memory, idea discovery, experiment automation, proof audits, and paper writing; also available as a standalone CLI.	Tools: Claude Code / Codex CLI skills and plugins, ARIS-Code CLI, MCP, Zotero, Obsidian LLM APIs: Anthropic Claude, OpenAI Codex; alternative executors/reviewers via OpenAI-compatible APIs
RD-Agent	Microsoft. Research and development agents for data science, quant factor/model evolution, Kaggle automation, paper-to-code implementation, and LLM fine-tuning.	Tools: Custom + LiteLLM, Docker, Qlib LLM APIs: OpenAI, Azure OpenAI, DeepSeek; LiteLLM-supported providers
AI-Scientist	Template-based scientific discovery system that automates idea generation, coding, experiments, manuscript writing, and paper review.	Tools: Custom (templates, LaTeX pipeline) LLM APIs: OpenAI, Anthropic Claude, DeepSeek, Gemini, OpenRouter, open-weight models
AutoResearchClaw	23-stage autonomous or human-in-the-loop research pipeline: idea → literature → domain-specific experiments → multi-agent review → LaTeX paper, with six intervention modes.	Tools: Custom pipeline, OpenClaw / ACP, Docker, LaTeX, OpenAlex, Semantic Scholar LLM APIs: OpenAI-compatible APIs (OpenRouter, DeepSeek, MiniMax); ACP-compatible coding agents
AI-Scientist-v2	Template-free scientific discovery system using progressive agentic tree search to generate hypotheses, run experiments, analyze results, and write manuscripts.	Tools: Custom (BFTS agentic tree search, AIDE) LLM APIs: OpenAI, Anthropic Claude (AWS Bedrock), Gemini
Agent Laboratory	End-to-end autonomous research workflow with specialized agents for literature review, experimentation, and report writing.	Tools: Custom multi-agent (arXiv, HuggingFace, LaTeX) LLM APIs: OpenAI (o1/o3/GPT-4o), DeepSeek
AI-Researcher	NeurIPS 2025 Spotlight. Fully autonomous system covering literature review, hypothesis generation, algorithm implementation, and manuscript preparation.	Tools: Custom + LiteLLM, Docker, Gradio LLM APIs: OpenAI, Anthropic, Gemini, DeepSeek, OpenRouter; LiteLLM-supported providers
claude-scholar	Semi-automated academic research assistant covering literature review, coding, experiments, writing, publication, and project knowledge management, with separate coding-agent branches.	Tools: Claude Code / Codex CLI / Kimi Code CLI / OpenCode, Zotero MCP, Obsidian, LaTeX LLM APIs: Anthropic Claude, OpenAI (Codex), Moonshot Kimi; OpenCode-configured providers
EvoScientist	Self-evolving research assistant with six sub-agents for planning, research, coding, debugging, analysis, and writing; persistent memory and reusable skills across CLI/TUI, WebUI, and messaging channels.	Tools: LangChain + LangGraph + DeepAgents, MCP, Docker LLM APIs: Anthropic, OpenAI, Google Gemini, DeepSeek, MiniMax, NVIDIA NIM, OpenRouter; configurable providers
Biomni	Stanford. General-purpose biomedical AI agent that autonomously executes research tasks across biology and medicine, combining LLM reasoning, retrieval, and tool/code use.	Tools: Custom biomedical agent + code execution, datalake, know-how library LLM APIs: Anthropic, OpenAI, Azure OpenAI, Gemini, Groq, AWS Bedrock, Ollama; custom OpenAI-compatible APIs
DeepScientist	Local-first autonomous research studio. Findings Memory + Bayesian optimization orchestrate baseline reproduction → branched experiments → LaTeX paper drafts.	Tools: Custom (Bayesian optimization, Findings Memory, Research Map), Git worktrees, LaTeX LLM APIs: Codex CLI, Claude Code, Kimi Code, OpenCode
DATAGEN	AI-driven multi-agent research assistant automating hypothesis generation, data analysis, visualization, and report writing.	Tools: LangChain + LangGraph, MCP, reusable skills, Firecrawl / fastCRW LLM APIs: OpenAI, Anthropic, Gemini, Ollama, Groq, Atlas Cloud, OrcaRouter
AutoSci	Memory-centric research system spanning literature review, ideation, experiments, analysis, and writing, with guided method iteration and separate Codex/OpenCode preview branches.	Tools: Memory-centric agent framework, persistent knowledge graph, web dashboard LLM APIs: Claude Code; preview support for Codex and OpenCode
InternAgent	Shanghai AI Lab. InternAgent-1.5 coordinates hypothesis generation, experiments, scientific paper reproduction, and deep research, with persistent memory across discovery runs.	Tools: Custom discovery / deep-research pipeline, coding-agent experiment backends, persistent memory LLM APIs: OpenAI-compatible APIs, OpenRouter, Anthropic Claude
Idea2Paper	Its Idea2Story module builds a literature knowledge graph, retrieves research patterns, and refines raw ideas into structured research stories through anchored multi-agent review.	Tools: Custom multi-agent pipeline, knowledge graph, embedding retrieval, anchored review LLM APIs: Configurable LLM endpoints and OpenAI-compatible embeddings APIs
NanoResearch	End-to-end idea-to-paper pipeline with real local or SLURM experiments; Evo mode adapts reusable skills, memory, and routing across research runs.	Tools: Nine-stage pipeline, Evo skills/memory/policy loop, local/SLURM execution, LaTeX LLM APIs: OpenAI-compatible APIs; Claude Code
K-Dense BYOK	Local AI co-scientist for literature, data analysis, code execution, figures, and reports, with 181 scientific skills and a hash-chained lab notebook that records executed steps and artifacts.	Tools: Local-first workspace, Pi agent runtime, scientific skills, specialist agents, MCP, Ollama LLM APIs: OpenRouter; direct OpenAI, Anthropic, Gemini and other provider APIs; supported subscriptions; Ollama / OpenAI-compatible local servers
data-to-paper	Turns research datasets into transparent, traceable, and verifiable manuscripts, with agents for analysis, interpretation, literature search, and writing.	Tools: Multi-agent pipeline, code execution, LaTeX, Semantic Scholar LLM APIs: OpenAI API; optional DeepInfra
Robin	FutureHouse's multi-agent system for scientific discovery, coordinating literature research, data analysis, and experimental planning.	Tools: Multi-agent scientific discovery system, LiteLLM, Edison platform, Docker / Jupyter LLM APIs: LiteLLM-supported providers; Edison platform for literature and data-analysis agents

↑ Back to contents

Projects focused on automated information gathering, literature review, and report generation.

Project & Stars	Description	Framework & LLM APIs
DeerFlow	ByteDance. DeerFlow 2.0 super-agent harness for research, coding, and reports, with sub-agents, persistent memory, reusable skills, project workspaces, and scheduled tasks.	Tools: LangChain + LangGraph, MCP, extensible skills, sandboxes, InfoQuest LLM APIs: OpenAI-compatible APIs, OpenRouter, vLLM; Codex CLI and Claude Code-backed providers
STORM	Stanford. LLM-powered knowledge curation system that generates full-length Wikipedia-like articles with citations. Features Co-STORM.	Tools: DSPy + LiteLLM, Streamlit LLM APIs: LiteLLM-supported providers; web search and custom document retrievers
GPT Researcher	Autonomous web and local-document research agent with source-tracked reports, PDF/Word/Markdown export, multi-agent workflows, and a coding-agent skill.	Tools: LangGraph / AG2, MCP, Claude Skill, FastAPI, NextJS LLM APIs: OpenAI, Anthropic Claude, Gemini; any OpenAI-compatible API
Tongyi DeepResearch	Alibaba. Open-weight Tongyi-DeepResearch-30B-A3B model for long-horizon information seeking, with ReAct and IterResearch-based heavy inference modes.	Tools: Custom (ReAct, IterResearch, GRPO RL); Serper, Jina, SandboxFusion LLM APIs: OpenAI-compatible, OpenRouter; Tongyi-30B-A3B, Dashscope/Bailian
Open Deep Research	LangChain. Open-source deep research framework with configurable MCP tools and search APIs.	Tools: LangChain + LangGraph, MCP, LangSmith LLM APIs: LangChain-supported providers (OpenAI, Anthropic, OpenRouter, Ollama); tool calling and structured output required
PaperQA2	Agentic RAG for scientific literature and local documents, with iterative retrieval, contextual summarization, citations, and support for PDF text, tables, figures, and equations.	Tools: Custom agentic RAG + LiteLLM, Pydantic, tantivy, multimodal document readers LLM APIs: OpenAI, Anthropic, Gemini, Ollama, llama.cpp; any LiteLLM provider
local-deep-research	Local-first agentic research with dynamically selected web and academic search engines, a searchable document library, MCP integration, and per-user encrypted storage.	Tools: LangChain + LangGraph, FastAPI, FAISS, SQLCipher, SearXNG LLM APIs: Ollama, LM Studio, llama.cpp; OpenAI, Anthropic, Gemini, OpenRouter, Requesty; compatible endpoints
Auto-Deep-Research	AutoAgent-based deep research assistant with browser and file workflows, configurable LiteLLM backends, and support for both function-calling and non-function-calling models.	Tools: AutoAgent Framework + LiteLLM, Docker LLM APIs: Anthropic, OpenAI, Gemini, Mistral, Groq, OpenRouter, DeepSeek; any OpenAI-compatible
OpenScholar	Retrieval-augmented scientific question answering with self-reflective generation and source-grounded citations; releases inference, retriever, and OpenScholar-8B training code.	Tools: Custom RAG, HuggingFace / PyTorch, torchtune training, retriever and reranker LLM APIs: OpenAI (GPT-4o), Llama 3.1 8B (self-hosted); Semantic Scholar API, You.com
OpenResearcher	Open deep-research training and inference recipe releasing the OpenResearcher-30B-A3B model, 96K agent trajectories, browser tools, and benchmark evaluation code.	Tools: Megatron-LM (training), vLLM (serving), HuggingFace, Tevatron, BM25 + Qwen3-Embedding, Serper LLM APIs: OpenResearcher-30B-A3B (open-weight release); OpenAI API (scoring)

↑ Back to contents

Coding, orchestration, and experiment-optimization systems that implement research ideas and run iterative empirical loops.

Project & Stars	Description	Framework & LLM APIs
OpenHands	Self-hosted developer control center for coding agents and automations. Agent Canvas runs agents across local, remote, and cloud backends and schedules reusable workflows.	Tools: Agent Canvas, OpenHands Agent Server / SDK, ACP, Docker LLM APIs: Configurable LLM providers; Claude Code, Codex, Gemini, and ACP-compatible agents
Aider	AI pair programming in your terminal. Uses repository maps for code context, edits multiple files, and integrates with Git, linting, and tests.	Tools: AI pair-programming CLI, LiteLLM, repository maps, Git integration LLM APIs: Anthropic, OpenAI, Gemini, DeepSeek, OpenRouter, Ollama; OpenAI-compatible APIs
SWE-agent	Princeton and Stanford. Configurable coding agent that edits repositories, uses tools, and fixes GitHub issues. Current development primarily focuses on mini-swe-agent.	Tools: YAML-configurable agent, LiteLLM, SWE-ReX LLM APIs: OpenAI, Anthropic; LiteLLM-supported providers
DeepCode	Coding agent for research reproduction and software engineering. Its Paper2Code workflow carries papers and reference repositories through planning, implementation, and experiment-based verification, with resumable sessions and parallel agents.	Tools: Custom agent harness, shared local service, TUI / Desktop / Web, Skills, MCP, Git worktrees LLM APIs: OpenAI, Anthropic, Gemini, DeepSeek, OpenRouter, Ollama, vLLM; OpenAI-compatible APIs
ClawTeam	Multi-agent orchestration for parallel experiments and coding tasks. CLI agents coordinate through shared task state and messages, while Git worktrees isolate changes; includes an autoresearch swarm example.	Tools: Python CLI, tmux, Git worktrees, shared tasks, file / ZeroMQ transport LLM APIs: Agent-dependent: Claude Code, Codex, OpenClaw, Cursor, Gemini CLI, and other CLI agents
Paper2Code	ICLR 2026. Converts machine-learning papers into runnable codebases through planning, analysis, and generation agents.	Tools: Multi-agent planning, analysis, and code generation; vLLM LLM APIs: OpenAI; open-weight models through vLLM
Paper2Agent	Turns research papers and their codebases into MCP servers and reusable skills, exposing scientific methods, data, and workflows as callable tools.	Tools: Agent Skills, parallel specialist agents, MCP, paper / codebase analysis LLM APIs: Agent-dependent: Claude Code, Codex, and other skill-compatible coding agents
DeepResearchAgent	Skywork. Composable self-evolving agent runtime that registers agents, tools, environments, and memory as versioned resources and tracks proposed, assessed, and committed improvements.	Tools: Custom (RSPL / SEPL self-evolution protocol), MMEngine configs, persistent memory, versioned resources LLM APIs: OpenRouter, OpenAI, Anthropic, Google Gemini; configurable API bases
MLE-agent	ML engineering agent for baseline construction, local execution, debugging, and Kaggle workflows, with code retrieval and project reports.	Tools: Python CLI, Kaggle, code RAG, local execution LLM APIs: OpenAI, Anthropic, Gemini, DeepSeek, Mistral, Ollama, vLLM; optional LiteLLM
AIDE	Open-source reference implementation of AIDE. Searches a tree of candidate ML programs, using execution results and a user-defined metric to guide drafting, debugging, and improvement. [paper]	Tools: Agentic tree search, Python CLI, Streamlit, Docker LLM APIs: OpenAI, Anthropic, Gemini; OpenAI-compatible local models such as Ollama
CORAL	Infrastructure for autoresearch-style experiment loops. Parallel coding agents share findings and reusable skills, while graders evaluate attempts; supports multi-island exploration and migration.	Tools: Multi-agent Git worktrees, graders, shared state, Docker, optional LiteLLM LLM APIs: Claude Code, OpenCode, Codex, Cursor Agent, Kiro; optional LiteLLM gateway

↑ Back to contents

Tools focused on reading, polishing, illustrating, presenting, and responding to reviews for scientific work.

Project & Stars	Description	Framework & LLM APIs
ChatPaper	Uses ChatGPT to summarize arXiv and local PDF papers, translate and polish manuscripts, analyze reviews, and draft reviewer responses.	Tools: PyMuPDF, arxiv.py, Flask, Gradio, Docker LLM APIs: OpenAI
PaperBanana	Reference-driven academic illustration framework with Retriever, Planner, Stylist, Visualizer, and Critic agents. Iteratively generates diagrams and plots from scientific content, with configurable vision-language and image models.	Tools: Five-agent illustration pipeline, Gradio, Streamlit, OpenRouter LLM APIs: Google Gemini; OpenAI, Anthropic, and other providers through OpenRouter
Paper2Poster	NeurIPS 2025. Converts a paper PDF into an editable academic poster (.pptx). Also provides a lightweight coding-agent skill for preparing poster-ready content.	Tools: Multi-agent pipeline, editable PowerPoint output, vLLM, Docker LLM APIs: OpenAI; open-weight language and vision models through vLLM
ChatReviewer	Uses ChatGPT to summarize paper strengths and weaknesses and suggest improvements. Its ChatResponse tool drafts point-by-point responses to reviewer comments; based on ChatPaper.	Tools: Python, PyMuPDF, tiktoken, Gradio, Docker, HuggingFace Spaces LLM APIs: OpenAI

↑ Back to contents

Reusable skill sets and plugin ecosystems that integrate with coding agents (Claude Code, Codex, Gemini CLI, etc.) to enable research workflows.

Project & Stars	Description	Framework & LLM APIs
scientific-agent-skills	177 scientific and research skills covering biology, chemistry, medicine, physics, engineering, Earth science, data analysis, and scientific communication, with reusable guidance for packages, databases, and workflows.	Tools: Agent Skills / Agent Plugins, scientific Python libraries, scientific databases LLM APIs: Agent-agnostic: Claude Code, Cursor, Codex, Google Antigravity, and compatible hosts
AI-Research-SKILLs	98 skills across 23 categories for the AI research lifecycle, including an autoresearch orchestration layer, ideation, training, evaluation, experiments, and paper writing.	Tools: Agent Skills, DeepSpeed, vLLM, LangChain, W&B, MLflow LLM APIs: Agent-agnostic: Claude Code, Codex, Gemini CLI, OpenCode, and compatible hosts
OpenClaw-Medical-Skills	869 medical and scientific skills spanning clinical reports, genomics, drug discovery, bioinformatics, structural biology, and biomedical databases, aggregated from multiple skill collections.	Tools: Agent Skills, OpenClaw / NanoClaw, biomedical tools and databases LLM APIs: Host-agent providers through OpenClaw / NanoClaw

↑ Back to contents

Curated collections and survey papers on the auto-research landscape.

Project & Stars	Description
awesome-autoresearch	Curated index of autonomous improvement loops and research agents inspired by Karpathy's autoresearch, including platform ports, domain adaptations, evaluations, and case studies.
awesome-ai-for-science	Curated list of AI tools, libraries, papers, datasets, and frameworks for scientific discovery across physics, chemistry, biology, and materials.
Autonomous-Agents	Daily-updated curated collection of research papers on autonomous LLM agents. Covers multi-agent systems, scientific computing, robotics, and more.
Awesome-Deep-Research	Curated collection of agentic deep research products, open-source implementations, research papers, benchmarks, and applications; includes a position paper on reasoning-driven search.

↑ Back to contents

This list focuses specifically on automating the scientific research process — not general-purpose AI agents. We include projects that target one or more stages of the research lifecycle:

📖 Literature Review → 💡 Idea Generation → 🔍 Novelty Check → 📐 Experiment Design →
💻 Code Implementation → 🚀 Experiment Execution → 📊 Result Analysis → ✍️ Paper Writing → 📝 Peer Review

General-purpose coding agents (OpenHands, Aider, SWE-agent) are included because they serve as critical infrastructure for the experiment execution stage.

PRs welcome! Please ensure the project:

• Has 500+ GitHub stars
• Is directly related to automating scientific research
• Is open-source with an active repository

Please keep entries sorted by star count (descending) within each category.

View the interactive Star History chart

CC0 1.0 Universal