AutoResearchClaw
Turns a research topic into experiments, verified citations, and paper-ready LaTeX deliverables.
Per-dimension scores and reasoning
Evidence shows: README describes sandbox isolation (Docker network policy), user confirmation gates, rollback on rejection, cost guardrails, and human-in-the-loop modes. However, least privilege principle is not explicitly stated, data flow transparency is limited, sensitive data handling (e.g., API keys) is only mentioned via environment variables, and dependency security is not addressed. Deductions: least privilege not explicit, data flow opaque, sensitive data handling details insufficient, dependency security unverified.
Evidence shows: README claims 2699 tests passed, but no test files are provided; an e2e_docker_sandbox.py script exists but is not part of the test suite. Dependency availability: pyproject.toml lists dependencies but no lock file or version ranges. Failure messages: README describes self-healing and error diagnosis, but no concrete error message examples. Deductions: test claims unverified, dependency locking missing, failure message examples insufficient.
Evidence shows: README describes multiple usage scenarios (CLI, OpenClaw, ACP agents), multiple intervention modes, cross-platform support. Capability boundaries: 23-stage pipeline and gates are described, but limitations are not explicit. Trigger precision: command examples and config are provided, but trigger conditions are not detailed. Environment fit: hardware detection and Docker sandbox are mentioned, but detailed environment requirements are not provided. Deductions: capability boundaries and trigger conditions not precise enough.
Evidence shows: README is well-structured, includes installation instructions, examples, changelog (News section), and MIT license. Naming stability: version numbers (v0.5.0) and release tags are provided. Known limitations: not explicitly listed. Maintenance responsibility: maintainers or contribution guidelines are not clearly stated. Deductions: known limitations missing, maintenance responsibility unclear.
Evidence shows: README describes outputs (paper, LaTeX, citations, etc.) and features (multi-source literature, verification, etc.), but no actual output examples or user feedback are provided. Marginal value: claims autonomous research, but no comparison with existing tools. Cost-benefit: cost guardrails are mentioned, but no cost data. Deductions: lack of actual output evidence and cost data.
Evidence shows: README claims an arXiv paper and ARC-Bench dataset, but no verification method is provided. Claims 2699 tests passed, but no test results are provided. Fact-inference separation: README distinguishes feature descriptions from claims, but does not explicitly label what is verified. Deductions: claims lack verifiable evidence, fact and inference not clearly separated.
- README claims 2699 tests passed, but no test files or results are provided, making it unverifiable.
- Dependencies are not pinned, which may affect reproducibility.
- Known limitations are not explicitly listed, which may mislead users about capability boundaries.
- Maintenance responsibility is unclear, and the update path is not transparent.
What does this agent do, and when should you use it?
AutoResearchClaw is a 23-stage research pipeline that takes a topic through literature work, experiments, analysis, paper drafting, and export. It runs from a CLI, integrates with OpenClaw, and can use ACP-compatible coding agents such as Claude Code, Codex CLI, Copilot CLI, Gemini CLI, OpenCode, and Kimi CLI as backends. Its phases cover scoping, literature discovery, synthesis, experiment design and execution, research decisions, writing, and finalization; Stage 15 can choose PROCEED, REFINE, or PIVOT. The documented outputs include Markdown and LaTeX papers, BibTeX, a citation-verification report, experiment code and metrics, charts, peer reviews, archived knowledge, and an Overleaf-ready deliverables folder. It fits research teams that can configure a model provider, run local sandbox or Docker-based experiments, and review generated research artifacts.
A run starts with researchclaw run --topic "..." --auto-approve or a collaborative mode. The pipeline decomposes the topic, searches OpenAlex, Semantic Scholar, and arXiv, then deduplicates and screens papers, extracts knowledge, generates and debates hypotheses, designs experiments, and generates hardware-aware Python code. The execution stages run experiments, detect NaN/Inf and runtime failures, and can perform targeted repair and iterative refinement; after result analysis, Stage 15 can send the run back to Stage 13 for REFINE or Stage 8 for PIVOT. The writing stages create an outline, draft, and multi-agent peer review, then export paper.tex, references.bib, verification_report.json, charts, experiment results, and deliverables/. Citation checking is documented as arXiv ID validation, CrossRef/DataCite DOI checks, Semantic Scholar title matching, and LLM relevance scoring; an optional MetaClaw bridge turns cross-run failures and warnings into reusable skills.
- An ML researcher wants to turn a testable topic into a draft with baselines, experiment code, statistical results, and a NeurIPS, ICLR, or ICML LaTeX target.
- A research group needs to search OpenAlex, Semantic Scholar, and arXiv while checking whether cited references are complete and relevant before writing.
- A faculty member or research engineer wants to steer hypothesis selection, experiment design, and paper drafting without manually operating every pipeline stage.
- A user working on CUDA, Apple MPS, or CPU-only hardware needs generated experiments to adapt to the detected local environment.
- A team conducting high-energy physics, biology, statistics, or generic Docker experiments needs execution routed by research domain.
What are this agent's strengths and limitations?
- The documented 23-stage workflow covers literature collection, experiments, and paper export rather than only generating prose.
- It combines OpenAlex, Semantic Scholar, and arXiv with a four-layer citation-verification flow designed to remove hallucinated references.
- Six intervention styles support fully automatic operation through stage-by-stage and collaborative review.
- It offers standalone CLI and Python API use alongside ACP-compatible coding agents and the OpenClaw bridge.
- Experiment execution includes hardware detection, sandbox or Docker configuration, NaN/Inf fast-fail behavior, and iterative repair.
- Adoption requires configuring an LLM provider or an ACP coding agent; the OpenAI-compatible route needs an endpoint, key environment variable, and model configuration.
- Literature retrieval, external APIs, and optional web search require network access; higher Semantic Scholar rate limits require an optional API key.
- Results depend on generated code, available hardware, and sandbox, Docker, or remote SSH setup; there is no documented guarantee of valid findings for every topic.
- Automated export does not remove the need for review: quality gates and some modes require approval, and the default security configuration blocks publishing without approval.
- Optional integrations such as MetaClaw, the OpenClaw bridge, specialist executors, and OpenCode add configuration and deployment complexity.
How do you install or deploy this agent?
Python 3.11+ is required. Run: git clone https://github.com/aiming-lab/AutoResearchClaw.git, cd AutoResearchClaw, python3 -m venv .venv && source .venv/bin/activate, pip install -e ., then researchclaw setup and researchclaw init. Initialization creates configuration; alternatively copy config.researchclaw.example.yaml to config.arc.yaml. For an OpenAI-compatible endpoint, set llm.base_url, llm.api_key_env, and llm.primary_model, then for example run export OPENAI_API_KEY="sk-...". The documented setup checks Docker and LaTeX; whether they are needed depends on the configured experiment and export workflow.
How do you use this agent?
For a first run: researchclaw run --config config.arc.yaml --topic "Your research idea" --auto-approve. For collaboration at key decisions: researchclaw run --topic "Your research idea" --mode co-pilot; reconnect to a paused run with researchclaw attach artifacts/rc-2026-xxx and use researchclaw approve, researchclaw reject, or researchclaw guide for stage decisions. For an ACP backend, set llm.provider to acp and configure llm.acp.agent, such as claude; the documentation also lists codex and other compatible commands. Completed outputs are placed in artifacts/rc-YYYYMMDD-HHMMSS-<hash>/deliverables/.
FAQ
Will it reliably produce a valid novel research result?
Must I use the OpenAI API?
Can I control automation and spending?
cost_budget_usd can be configured; the pipeline is documented to pause when the configured budget is exceeded.