Biomedical AI Agents in 2026: 8 Open-Source Tools Compared

May 10, 2026

What is a biomedical AI agent?

A biomedical AI agent is an autonomous LLM-based system that performs biomedical research tasks end-to-end — literature search, sequence analysis, single-cell processing, experimental design, or systematic review — instead of acting as a chat assistant. Eight open-source projects now cover the major use cases. This post compares them and tells you which one to pick for which job.

All eight are MIT/Apache-licensed, pre-2.0, and actively maintained or paper-anchored. Star counts as of 2026-05-10.


The 8 projects at a glance

ProjectStarsLicenseBest forSource
Biomni2,900Apache-2.0General biomedical platformStanford SNAP Lab
MedgeClaw856MITClinical / medical researchxjtulyc
ClawBio678MITLocal-first bioinformaticsClawBio org
BioClaw323MITCommunity bioinformaticsRunchuan-BU
CellVoyager202MITSingle-cell RNA-seqStanford Zou Group
BioMaster91MITMulti-agent bioinformaticsai4nucleome
BioMedAgent57MITPaper-anchored, peer-reviewedIndependent
Darwin1MITSystematic review + clinical evidenceyejunbin

Star count alone is misleading — Biomni's 2.9K reflects general-purpose breadth, while CellVoyager's 202 reflects depth in one task (scRNA-seq) where it has a Nature Methods paper. Use the comparison sections below to pick by job.


Which biomedical AI agent should I use?

Use this decision tree:

  • General biomedical research, broad coverage → Biomni
  • Clinical evidence, systematic reviews → Darwin
  • Local execution, no data leaves your machine → ClawBio
  • Single-cell RNA-seq analysis → CellVoyager
  • Genetic perturbation experimental design → BioDiscoveryAgent
  • Real-time dashboard + JupyterLab → MedgeClaw
  • Multi-agent coordination on bioinformatics tasks → BioMaster
  • Paper-anchored, peer-reviewed reference implementation → BioMedAgent

Now in detail.


Biomni — the general-purpose platform

Biomni is Stanford SNAP Lab's general-purpose biomedical AI agent, the highest-starred open-source biomedical agent on GitHub (2,900⭐ as of 2026-05-10). It targets broad coverage across biomedical tasks rather than depth in any single workflow, with a unified agent architecture that accepts diverse research queries.

The headline trade-off: breadth. If your job spans literature search, sequence analysis, and pathway enrichment in a single session, Biomni is the safe default. If you need a specific workflow optimised end-to-end (e.g. scRNA-seq), one of the specialists below will outperform it.

Best for: lab teams wanting a single tool to standardise on, before they know which sub-tasks dominate their day.


MedgeClaw — clinical research with a live dashboard

MedgeClaw is a clinical/medical research assistant built on Claude Code, bundling 140 K-Dense scientific skills with a real-time research dashboard, RStudio (port 8787), JupyterLab (port 8888), and Feishu rich-card output. 856 stars as of 2026-05-10.

The differentiator is operational visibility. Most agents give you a final answer; MedgeClaw shows you what's running, lets you intervene mid-pipeline through RStudio/JupyterLab, and surfaces results into Feishu — useful for Chinese clinical research teams that already use Feishu for collaboration.

Best for: clinical research teams that need observability and want to keep statistical work in R/Python notebooks instead of letting the agent black-box it.


ClawBio — local-first bioinformatics

ClawBio is the first bioinformatics-native AI agent skill library, built on OpenClaw with a local-first, reproducible architecture. 678 stars. The design choice that defines it: data never leaves your machine. Genomic data, patient samples, equity-focused workflows — all run in-process.

Coverage spans genomics, population genetics, and bioinformatics workflows where data sovereignty matters (academic medical centres, clinical labs, jurisdictions with strict patient data laws).

Best for: institutions that cannot use cloud-hosted agents for compliance or jurisdictional reasons. Also the right choice when reproducibility audits are part of your pipeline (every run produces a verifiable artefact).


BioClaw (Runchuan-BU) — community bioinformatics

BioClaw is an active community-driven bioinformatics agent, distinct from ClawBio despite the similar name (they are different teams). 323 stars. Less infrastructurally opinionated than ClawBio — closer to a flexible toolkit you wire into your existing workflow.

Best for: researchers who want to compose bioinformatics workflows themselves rather than adopt a packaged platform. The lower opinionation also makes it easier to extend with custom skills.

Worth knowing: there are at least four projects with the BioClaw / ClawBio name family. See our BioClaw vs ClawBio comparison for the disambiguation.


CellVoyager — Nature Methods scRNA-seq specialist

CellVoyager is an autonomous single-cell RNA-seq analysis agent from Stanford's Zou Group, published in Nature Methods (s41592-026-03029-6). 202 stars. Self-debugging code generation plus hypothesis generation are its differentiators — it doesn't just run the standard pipeline, it proposes follow-up analyses based on what the data shows.

Best for: scRNA-seq labs specifically. If your work is single-cell, the Nature Methods publication and self-debugging behaviour together make it hard to beat — the depth advantage over a general-purpose agent is real.

Trade-off: outside scRNA-seq, you'd be using a single-cell-tuned tool for the wrong job.


BioDiscoveryAgent — Stanford experimental design

BioDiscoveryAgent is a closed-loop genetic-perturbation experiment-design agent from Stanford SNAP. 102 stars, paper-anchored. The novelty is the loop: agent proposes experiments → you (or simulated lab) run them → agent updates beliefs and proposes the next round.

Best for: labs running CRISPR screens or perturbation studies who want to optimise experimental sequence rather than run all combinations. Also valuable as a reference implementation for closed-loop biological discovery — the architecture transfers to other experimental design problems.


BioMaster — multi-agent coordination

BioMaster is a multi-agent system for bioinformatics tasks, with specialised agents that coordinate across data preparation, analysis, and interpretation. 91 stars, from the ai4nucleome group (same authors as POPGENAGENT).

Best for: tasks where decomposition matters — e.g. a complex pipeline that splits into ingestion, QC, analysis, and writeup, where having dedicated agents for each phase reduces context bloat compared to a single monolithic agent.

The architecture is the lesson here even if you don't use BioMaster directly: multi-agent decomposition is the answer to long-context degradation in biomedical workflows.


BioMedAgent — Nature BME peer-reviewed reference

BioMedAgent is published in Nature Biomedical Engineering (s41551-026-01634-6). 57 stars. We've marked it stable rather than dormant — paper-locked projects are intentionally low-activity because the paper is the deliverable, not ongoing code.

Best for: researchers who need a peer-reviewed citation for "AI-assisted biomedical research" in their methods section. The codebase is reference-quality but not actively maintained — fork it, don't expect upstream updates.


Darwin — PRISMA-driven systematic review

Darwin is a biomedical research agent forked from feynman, specialised for clinical evidence work (added to Claw4Science 2026-05-09). 1 star at time of writing. The differentiation is methodological: PRISMA-compliant systematic reviews, clinical evidence grading (RCT > cohort > preprint), MIQE/CONSORT/ARRIVE-compliant protocols, integrated BLAST/GEO/AlphaFold, and standard nomenclature throughout (HGNC, RRID, UniProt, ChEMBL).

Best for: researchers writing systematic reviews or meta-analyses for clinical journals. The compliance-framework grounding is unusual — most AI agents don't enforce reporting standards. Also useful for pharmacovigilance work where evidence hierarchy matters.

Read the full deep dive on the project's repo — early-stage but methodologically serious.


Approach comparison

The eight projects sit on three orthogonal axes:

AxisExamples
Generalist vs specialistBiomni (generalist) ←→ CellVoyager (scRNA-seq specialist)
Cloud vs localMedgeClaw (cloud-friendly) ←→ ClawBio (local-first)
Code-first vs evidence-firstBioMaster (pipelines) ←→ Darwin (PRISMA reviews)

Most directories rank biomedical agents by stars or recency. We rank by fit — pick the one whose default assumptions match your job, then evaluate quality.


FAQ

Are these biomedical AI agents free?

Yes. All eight are open-source under MIT or Apache 2.0. Free for personal, academic, and commercial use including modification.

Do I need GPU access?

It depends on the underlying LLM. Cloud-hosted Claude / GPT-4 / Gemini work without GPUs. Local LM Studio / Ollama / vLLM setups need GPU memory matching your model size (typically 16-48 GB VRAM).

Which one is HIPAA-compliant?

ClawBio's local-first architecture is the easiest to reconcile with HIPAA — data never leaves the machine. Cloud-routed agents (Biomni, MedgeClaw) need a HIPAA-compliant LLM provider (Anthropic / OpenAI Enterprise / Azure OpenAI under BAA).

Can I use them for drug discovery specifically?

These eight focus on biomedical research broadly. For drug discovery / molecular work specifically, see our drug-molecular project group — DrugClaw, OriGene, ChemCrow, MDCrow, AIAgents4Pharma, ChemGraph, and others.

Which has the best peer-reviewed credentials?

CellVoyager (Nature Methods) and BioMedAgent (Nature Biomedical Engineering) are both peer-reviewed in top venues. Darwin grounds its methodology in published reporting standards (PRISMA, CONSORT) without itself being peer-reviewed yet.

How do I install all eight?

Each project has its own install path; we list them on the per-project pages. The fastest survey is to use Claw4Science's project grid — every entry has install instructions linked from the card.


Try it

If you build a biomedical AI agent that fits a use case we haven't covered, submit it — we feature serious biomedical work regardless of star count, as long as it has a clear use case and methodology.


Last updated: 2026-05-10. Star counts and links verified against GitHub on the publication date.