ScienceClaw is a science research AI agent. Say "分析 TP53 在肝癌中的作用" and it autonomously searches 15+ literature sources, queries 77+ databases, runs survival analysis in R, generates publication-quality figures, and delivers a full report with real citations — no fabrication, no hallucination.
Zero custom code. Built entirely on OpenClaw with one markdown file (SCIENCE.md, ~600 lines) and 266 domain skills. The model does 99% of the work; the markdown teaches it how to be a scientist.
Prompt: "Investigate the role and significance of THBS2 in tumors"
ScienceClaw autonomously searched PubMed, queried TCGA via cBioPortal and TIMER2.0, ran survival analyses in R, and compiled a 30-page report with 87 citations.
Key findings:
- THBS2 is significantly upregulated in 17 out of 33 TCGA cancer types
- Combined THBS2 + CA19-9 panel achieved diagnostic AUC of 0.96 in a retrospective pancreatic cancer cohort — but dropped to 0.69 in a prospective validation set
- Tumor microenvironment analysis revealed THBS2 correlation with M2 macrophage infiltration across multiple cancer types
Prompt: "Survey the applications of LLM in biomedicine"
ScienceClaw conducted a systematic literature search across PubMed, Semantic Scholar, and OpenAlex, then synthesized findings into a structured survey with trend analysis and visualizations.
Key findings:
- Medical LLM publications grew 570x in two years — from 8 in 2022 to 4,562 in 2024
- Med-PaLM 2 reached 86.5% accuracy on USMLE, surpassing the expert physician threshold
- The healthcare LLM market is projected to reach $110B by 2030
Prompt: "分析 TP53 在肝癌中的作用"
ScienceClaw auto-matches the gene-landscape Recipe and executes a 6-step pipeline autonomously: literature search → TCGA expression profiling → survival analysis → immune infiltration → pathway enrichment → structured report with METHODS.md.
All output files are saved to ~/.scienceclaw/workspace/projects/tp53-liver-cancer-<date>/ and can be exported with one command: /export word, /export pptx, or /export latex.
Browse all 6 Research Recipes →
git clone https://github.com/Zaoqu-Liu/ScienceClaw.git && cd ScienceClaw
bash scripts/setup.sh # installs deps, configures API key (interactive)
./scienceclaw run # starts gateway + opens TUI — doneChina users: setup will ask for an API key. Use DeepSeek — direct access, no proxy, ¥1/M tokens. Or use OpenRouter as a relay for all providers.
Prerequisites
| Requirement | Version | Notes |
|---|---|---|
| Node.js | >= 22 | nodejs.org |
| Python | >= 3.10 | For code execution (R, Julia optional) |
| Docker | Latest | Optional — sandboxed execution |
Troubleshooting
./scienceclaw models # which models work? diagnose 404/403/429
./scienceclaw doctor # full system health checkOne-shot mode & channels
./scienceclaw ask "分析 BRCA1 在乳腺癌中的作用" # one-shot, no TUI
./scienceclaw add telegram # or discord, slack, whatsapp, feishu, wechat
./scienceclaw channels # list configured channels| Capability | Details |
|---|---|
| Search literature | 15+ sources — PubMed, Semantic Scholar, OpenAlex, Europe PMC, and more |
| Query databases | 77+ databases — UniProt, PDB, NCBI, ChEMBL, STRING, GTEx, ClinicalTrials.gov, and more |
| Run code | Python, R, Julia via bash — install packages on the fly |
| Generate figures | Journal-spec palettes (NPG, Lancet, JCO, NEJM), publication-ready sizing |
| Write reports | Real citations from search results, never fabricated |
| Review research | 8-dimension ScholarEval rubric for systematic quality assessment |
| Research Recipes | 6 pre-built workflows — gene landscape, target validation, literature review, and more |
| Export deliverables | One command to Word, PowerPoint, or LaTeX from project results |
| Monitor literature | /watch tracks topics on PubMed, alerts on new papers at session start |
Six pre-built research workflows that execute complete multi-step analyses from a single prompt. ScienceClaw auto-detects which Recipe matches and runs the full pipeline autonomously.
| Recipe | Trigger Examples | What It Does |
|---|---|---|
| gene-landscape | "分析 TP53 在肝癌中的作用" | Literature → TCGA expression → survival → immune → pathway → report |
| target-validation | "评估 EGFR 的成药性" | Literature → STRING → ChEMBL → DrugBank → trials → patents → report |
| literature-review | "综述 CRISPR 在基因治疗中的应用" | Multi-source 50+ → filter → full text → trend chart → structured review |
| diff-expression | "分析这个表达矩阵" | QC → DESeq2/limma → volcano + heatmap → GO/KEGG → report |
| clinical-query | "NSCLC 的最新治疗方案" | ClinicalTrials → guidelines → drugs → summary table |
| person-research | "调研张三教授" | OpenAlex → PubMed → citations → themes → profile report |
./scienceclaw recipes # list all Recipes
./scienceclaw ask "分析 TP53 在肝癌中的作用" # auto-matches gene-landscape| Feature | Description |
|---|---|
| Research Recipes | 6 one-liner-to-full-workflow templates (see above) |
| Export to Word/PPT/LaTeX | /export word, /export pptx, /export latex — generate formatted deliverables from project results |
| Literature Monitoring | /watch TOPIC — track new publications on PubMed, alerts at session start |
| Research Memory | Structured findings stored in JSONL — cross-session, cross-project recall via /recall |
| METHODS.md | Auto-generated Methods section after deep analyses, ready for paper insertion |
| Smart Task Routing | Quick tasks (single lookup) stay in chat; deep tasks get project directories |
| Follow-up Suggestions | Data-driven next-step suggestions after every multi-step analysis |
| Session Greeting | Context-aware greeting — returning users see recent project status + pending alerts |
| First-run Welcome | Guided onboarding for new users with actionable examples |
CLI recipes / ask |
./scienceclaw recipes to browse, ./scienceclaw ask "..." for one-shot queries |
ScienceClaw inherits all channel integrations from OpenClaw. Connect your preferred interface:
| Channel | How to use |
|---|---|
| Terminal UI | scienceclaw tui |
| Web Dashboard | scienceclaw dashboard |
| Telegram | Setup guide |
| Discord | Setup guide |
| Slack | Setup guide |
| Feishu / Lark | Setup guide |
| Setup guide | |
| Setup guide | |
| Matrix | Setup guide |
| + more | scienceclaw channels --help |
ScienceClaw = OpenClaw engine + SCIENCE.md (~600 lines) + 266 Skills (markdown)
= 0 lines of custom code
No TypeScript. No Python servers. No MCP. No plugins. No middleware. The scienceclaw bash wrapper (~130 lines) manages the gateway lifecycle. Everything else is markdown that teaches the model how to be a scientist.
| Layer | Components |
|---|---|
| User | Terminal UI, Web Dashboard, Telegram, Discord, Slack, Feishu, WeChat, WhatsApp, Matrix |
| Gateway | OpenClaw gateway — routes messages, manages sessions, handles tool calls (port 18789) |
| Agent | SCIENCE.md (identity + research discipline) + 266 domain skills (loaded on demand) |
| Tools | web_search (Brave), bash (Python/R/Julia + curl to REST APIs) — two tools do everything |
When models get smarter, ScienceClaw gets smarter — no code to update, no integrations to fix. See Architecture docs for the full design rationale.
ScienceClaw searches across 15+ sources, cross-references results, and verifies citations before including them in reports.
| Category | Sources |
|---|---|
| Biomedical literature | PubMed, PubMed Central, Europe PMC |
| Broad academic | Semantic Scholar, OpenAlex, CrossRef, CORE |
| Preprints | bioRxiv, medRxiv, arXiv |
| Clinical | ClinicalTrials.gov, WHO ICTRP |
| Patents & grants | Google Patents, NIH RePORTER |
| General | Google Scholar, Web search |
77+ databases across 9 disciplines, all accessed through their public REST APIs via bash + curl.
| Discipline | Databases | Count |
|---|---|---|
| Genomics & Transcriptomics | NCBI Gene, Ensembl, UCSC Genome Browser, GEO, TCGA, GTEx, ENCODE | 10+ |
| Proteomics & Structure | UniProt, PDB, AlphaFold DB, InterPro, Pfam, SWISS-MODEL | 8+ |
| Pathways & Interactions | STRING, BioGRID, KEGG, Reactome, WikiPathways, IntAct | 8+ |
| Pharmacology & Drug Discovery | ChEMBL, DrugBank, PubChem, PharmGKB, DGIdb, TTD | 8+ |
| Disease & Phenotype | OMIM, DisGeNET, ClinVar, GWAS Catalog, HPO, Orphanet | 8+ |
| Immunology | IEDB, IMGT, ImmPort, TIMER2.0, TCIA | 6+ |
| Microbiome | GMrepo, gutMDisorder, BugBase, MicrobiomeDB | 5+ |
| Clinical & Epidemiology | ClinicalTrials.gov, GBD, WHO GHO, SEER, cBioPortal | 7+ |
| Model Organisms | MGI, FlyBase, WormBase, ZFIN, RGD, SGD | 7+ |
Each skill is a markdown file that teaches the model how to perform a specific analysis — complete with API patterns, code templates, and validation steps.
| Domain | Count | Skills |
|---|---|---|
| Bioinformatics | 30+ | scanpy, anndata, pydeseq2, arboreto, biopython, deeptools, pysam |
| Visualization | 35+ | matplotlib, seaborn, plotly, visualization, networkx |
| Drug Discovery | 20+ | chembl-database, rdkit, zinc-database, alphafold-database, adaptyv, medchem |
| Clinical & Survival | 15+ | clinicaltrials-database, scikit-survival, clinical, fda-database |
| Single-cell | 10+ | scanpy, scvi-tools, cellxgene-census, anndata |
| Genomics | 15+ | gene-database, ensembl-database, gwas-database, clinvar-database, geo-database |
| Databases | 20+ | uniprot-database, pdb-database, string-database, opentargets-database, reactome-database |
| Machine Learning | 10+ | scikit-learn, shap, aeon, statistics, exploratory-data-analysis |
| Scientific Writing | 15+ | academic-literature-search, writing, review-writing, peer-review, venue-templates |
./scienceclaw skills # browse all 266 skills by domain
./scienceclaw skills search "survival" # search by keywordLocal — already covered in Quick Start.
Docker — sandboxed Python/R/Julia execution:
docker compose -f docker/docker-compose.yml upSee Deployment Guide for production options including autostart, reverse proxy, and cloud deployment.
Contributions are welcome. Please read CONTRIBUTING.md before submitting a pull request.
LIU Zaoqu
International Academy of Phronesis Medicine (Guangdong) · π-HuB infrastructure
Contact: liuzaoqu@163.com
This project is licensed under the MIT License.
ScienceClaw — Your AI Research Colleague.