MedgeClaw is an open-source biomedical AI research assistant built on OpenClaw and Claude Code. You send a message via WhatsApp, Telegram, Slack, or Discord describing your analysis — RNA-seq differential expression, drug target screening, survival analysis, single-cell clustering, etc. — and MedgeClaw runs it autonomously using R and Python, then shows results in RStudio or JupyterLab. No command-line required.

What analyses can MedgeClaw run?

The 140 K-Dense Scientific Skills cover: genomics (DESeq2, edgeR, limma for differential expression; Seurat and Scanpy for single-cell), drug discovery (ChEMBL queries, target screening, pathway enrichment), clinical research (survival analysis with survminer, CHARLS/NHANES dataset workflows), literature research (PubMed/Semantic Scholar queries), and scientific visualization (publication-ready figures, ggplot2, plotly). Skills list which API keys they require.

What is the architecture?

Messages → OpenClaw gateway → biomed-dispatch skill → Claude Code execution layer → R + Python analysis environment → results in RStudio/JupyterLab and back to your chat. Claude Code runs K-Dense skills which are pre-written analysis templates for specific biomed workflows. OpenClaw handles the messaging routing; Claude Code handles the execution; R/Python do the actual computation.

What are K-Dense Scientific Skills?

K-Dense Scientific Skills are a library of 140 domain-specific Claude Code skills for scientific computing — pre-written analysis workflows for bioinformatics, drug discovery, clinical research, and ML. Each skill knows which packages to use, how to structure the analysis, what outputs to produce, and which API keys it needs. Think of them as the scientific equivalent of Claude Code's coding skills, but for research workflows.

How does it compare to using ChatGPT or Gemini for bioinformatics?

General LLMs know about bioinformatics conceptually but don't run actual analyses against your data — they generate code suggestions you then execute yourself. MedgeClaw executes the analysis end-to-end: it calls DESeq2 on your actual count matrix, runs the enrichment analysis, generates the heatmap, and returns the figure. The 140 K-Dense skills also encode domain-specific best practices that a general LLM prompt doesn't.

What model providers does MedgeClaw support?

Anthropic Claude (default, best quality), MiniMax 2.1, GLM-4.7 via Z.ai, DeepSeek, and Ollama for fully local offline inference. All providers use the same configuration — just change the base URL and model in .env. For non-Anthropic providers, set ANTHROPIC_SMALL_FAST_MODEL to avoid Claude Code hanging on pre-flight checks.

MedgeClaw: Biomedical AI Research Assistant Built on OpenClaw

By Prahlad Menon Published 2026-03-23 4 min read

The pattern of building specialized research tools on top of AI agent platforms is accelerating. gstack gave solo developers a virtual tech company. Claude Code Game Studios gave game developers a 48-agent studio hierarchy. Now MedgeClaw gives biomedical researchers an AI assistant that actually runs their analyses.

What makes MedgeClaw different: it’s built on OpenClaw — the same agent platform running this research and this blog. A researcher at Xi’an Jiaotong University took OpenClaw + Claude Code and turned it into a domain-specific research assistant for biomedicine. That’s the architecture worth understanding.

The Workflow

You (WhatsApp / Telegram / Slack / Discord)
  "Run differential expression analysis on my count matrix, 
   compare treatment vs control, filter FDR < 0.05"
        ↓
OpenClaw Gateway  (conversation + routing layer)
        ↓  biomed-dispatch skill
Claude Code  (execution layer)
        ↓  K-Dense Scientific Skills (140 domain workflows)
R + Python Analysis Environment
  DESeq2 / edgeR / Seurat / Scanpy / survminer / ...
        ↓                          ↓
Research Dashboard            RStudio / JupyterLab
  (real-time progress)         (results + interactive)
        ↓
Results back to your chat

No command line. No managing conda environments. No forgetting which function signature DESeq2 uses for multi-factor designs. You describe the analysis; Claude Code writes and executes it using a K-Dense skill that already knows the right approach.

The 140 K-Dense Scientific Skills

K-Dense skills are pre-written analysis workflows — the scientific equivalent of Claude Code’s coding skills, but for research. Each skill encodes domain-specific best practices: which packages, which statistical tests, which visualization conventions, what QC steps to run before the main analysis.

Genomics:

Differential expression (DESeq2, edgeR, limma)
Single-cell RNA-seq (Seurat pipeline: QC → normalization → clustering → annotation)
GSEA and pathway enrichment (clusterProfiler, fgsea)
Variant annotation

Drug Discovery:

ChEMBL target queries
Compound screening workflows
Target-pathway mapping
ADMET property prediction

Clinical Research:

Survival analysis (survminer, lifelines)
CHARLS/NHANES dataset workflows (pre-built cohort extraction)
Clinical variable selection
Epidemiological analysis

Scientific Visualization:

Publication-ready figures (ggplot2, matplotlib)
Heatmaps (pheatmap, ComplexHeatmap)
CJK font support for Chinese research outputs

The skills list which API keys they require — NCBI for literature queries, ChEMBL for drug data, Semantic Scholar for citation networks.

The Analysis Environment

The Docker container includes:

R stack: DESeq2, edgeR, limma, Seurat, clusterProfiler, fgsea, survival, survminer, ggplot2, pheatmap, ComplexHeatmap, GSVA, AnnotationHub

Python stack: Scanpy, anndata, scvi-tools, BioPython, PyDESeq2, lifelines, nibabel, pydicom, scikit-learn, pandas, numpy, scipy, matplotlib, seaborn, plotly

Interfaces: RStudio Server (browser-based R IDE), JupyterLab (Python + R kernels), real-time Research Dashboard

The environment runs in Docker so your main system stays clean. Data and outputs persist through Docker volumes.

Why OpenClaw?

The messaging layer is what makes this usable outside of a terminal session. OpenClaw handles WhatsApp, Telegram, Slack, Discord, and more — you can send an analysis request from your phone, check in on progress via chat, and receive results as rich messages or file attachments. The underlying computation happens on a server; you interact through whatever messaging app you’re already using.

The biomed-dispatch skill is the bridge — it routes your natural language request to the right K-Dense skill and manages the conversation context across multi-step analyses.

Model Flexibility

MedgeClaw isn’t locked to Claude. It supports Anthropic (default), MiniMax, GLM-4.7, DeepSeek, and Ollama for fully local offline inference. For research environments with data sovereignty requirements, the Ollama path means no data ever leaves your infrastructure.

The only configuration difference between providers is the base URL in .env. For non-Anthropic providers, set ANTHROPIC_SMALL_FAST_MODEL to a model your endpoint supports — Claude Code uses a lightweight model for pre-flight checks that most third-party proxies don’t expose.

The Broader Pattern

MedgeClaw is the third project this month building domain-specific agent systems on OpenClaw’s gateway layer. The pattern is consistent: OpenClaw handles the conversation and routing; Claude Code handles execution; a domain-specific skill library encodes the specialized knowledge.

For bioinformatics, that specialization matters. DESeq2 has specific requirements around count normalization and dispersion estimation that a general-purpose LLM prompt doesn’t reliably get right. Pre-written skills that encode those requirements — maintained by domain experts, versioned in git, auditable — are more reliable than ad-hoc prompts in every session.

The 34% fork rate (339 forks out of 982 stars as of launch month) suggests the research community is building on it, not just reading about it. That’s the signal worth watching.

→ github.com/xjtulyc/MedgeClaw · K-Dense Scientific Skills