AutoResearchClaw is an open-source autonomous research pipeline for academic research, with the slogan "Chat an Idea. Get a Paper.", meaning that users propose a research idea and the system automatically produces a complete paper. The project is developed by aiming-lab, uses the MIT license, and the related paper has been published on arXiv.
Purpose AutoResearchClaw aims to automate the research process from idea to paper. After users input a research topic, the system retrieves real literature, designs and executes experiments, performs statistical analysis, organizes multi-agent peer review, and ultimately generates a LaTeX paper for conferences such as NeurIPS/ICML/ICLR. It emphasizes not producing fabricated references, and self-repairs when experiments fail, pivots when hypotheses do not hold, and removes citations when they are fabricated.
Main capabilities
- A 23-stage, 8-phase pipeline covering the complete process from idea to paper.
- Literature sources include OpenAlex, Semantic Scholar, and arXiv.
- Hardware-aware sandbox experiments, with automatic detection of GPU/MPS/CPU.
- Multi-agent subsystems: CodeAgent, BenchmarkAgent, FigureAgent.
- Multi-agent peer review and 4 rounds of paper quality auditing, including AI trace detection, 7-dimensional review scoring, and the NeurIPS checklist.
- Output to artifacts/rc-YYYYMMDD-HHMMSS-<hash>/deliverables/, including compilable LaTeX, BibTeX, experiment code, and figures.
- Supports fully autonomous mode and Human-in-the-Loop copilot mode.
- Supports MetaClaw cross-run learning, turning failure experiences into reusable skills and injecting them into subsequent runs.
- Supports loading open-source and custom skills, with 20 built-in preset skills covering scientific writing, experimental design, chemistry, biology, and more.
- Supports multi-domain experimental agents: high-energy physics (ColliderAgent), biology (COBRApy metabolic modeling), statistics (simulation study agent), and a general Docker executor covering chemistry/materials.
- Provides the ARC-Bench benchmark, including 55 open autonomous research topics covering machine learning, high-energy physics, quantum, biology, and statistics.
Usage The recommended way is to use OpenClInstall and start a conversation in aw: OpenClaw automatically handles git clone, pip install, configuration setup, and pipeline execution. Users only need to send a message like "Research X". It can also be run independently via CLI, or connected to any ACP-compatible coding agent (Claude Code, Codex CLI, Copilot CLI, Gemini CLI, Kimi CLI, etc.), without an API key. Through the OpenClaw bridge, research can also be launched from messaging platforms such as Discord, Telegram, Lark (Feishu), and WeChat. The gated stages in the pipeline (5, 9, 20) can pause to wait for human approval, or use --auto-approve to approve automatically; if rejected, the pipeline rolls back. Co-pilot mode (--mode co-pilot) provides deep human-machine collaboration in stages 7-8 (Idea Workshop), stage 9 (Baseline Navigator), and stages 16-17 (Paper Co-Writer).
Use cases Suitable for researchers, research teams, and AI researchers who want to automate literature review, experimental design, code generation, experiment execution, statistical analysis, and paper writing. It is especially suitable for scenarios that require rapid exploration of research ideas, multi-domain experiments, or completing a first draft of a paper under human-machine collaboration.
Limitations The materials do not provide specific fees, prices, platform support scope, or performance guarantees. Some advanced features (such as MetaClaw integration and HITL mode) are disabled by default and need to be explicitly enabled. Domain-specific features such as HEP-ph physics mode require specific configuration conditions to be met. The system depends on an external LLM API or an ACP-compatible coding agent, and the specific available models and configurations should be referenced in the official documentation.
