Best Claude Plugins for Life-Science Research in 2026
Reviewed
Best Claude Plugins for Life-Science Research in 2026
The best Claude plugins for life-science research in 2026 are Anthropic's Bio Research plugin for preclinical R&D and K-Dense's Scientific Agent Skills for computational biology. Claude Science, launched in June, is a separate app rather than a plugin. Together they cover literature search, single-cell analysis, sequencing pipelines and target discovery. Neither plugin checks whether a retrieved paper actually supports the sentence it is cited for, which is the gap this post also covers.
Claude plugins, connectors and skills for life-science research: what each one is
Claude has three extension types, and life-science tools ship as all three. A connector links Claude to one external service through the Model Context Protocol (MCP), such as PubMed, Benchling or ChEMBL. A skill is a folder of instructions and scripts that teaches Claude a procedure, such as filtering a single-cell RNA-seq dataset. A plugin bundles connectors, skills, slash commands and subagents into one install.
Where they run matters. Connectors work in Claude.ai chat, Claude Desktop and Cowork. Plugins run in Cowork and Claude Code, and both of those need a paid Claude plan. A researcher who only uses Claude in the browser needs connectors. A researcher who runs multi-step analyses on local files benefits from plugins.
Most of the life-science ecosystem arrived in two waves. In October 2025, Anthropic launched Claude for Life Sciences with connectors for Benchling, BioRender, PubMed, Wiley Scholar Gateway, Synapse.org and 10x Genomics. In January 2026 it added Medidata, ClinicalTrials.gov, ToolUniverse, bioRxiv and medRxiv, Open Targets, ChEMBL and Owkin's Pathology Explorer. The plugins below are largely ways of packaging those connectors with the skills that make them useful.
Bio Research: Anthropic's official Claude plugin for preclinical life-science R&D
Bio Research is the plugin to install first. It is Anthropic-verified, open source, and published in Anthropic's knowledge-work-plugins collection. It bundles 10 MCP integrations with 6 analysis skills, so one install gives Claude the literature, target and trial databases most preclinical teams use.
The connectors cover four jobs:
- Literature: PubMed, bioRxiv and Wiley full-text access.
- Target and compound discovery: Open Targets and ChEMBL.
- Clinical context: ClinicalTrials.gov.
- Lab and data platforms: Benchling, Synapse, BioRender and Owkin.
The skills cover the following:
- Quality control for scRNA-seq data following scverse practice.
- scvi-tools for batch correction and cell-type annotation.
- nf-core pipelines for RNA-seq, variant calling and ATAC-seq on local or GEO/SRA data.
- Drafting clinical trial protocols.
- Converting output from more than 40 instrument types to the Allotrope Simple Model.
- A framework for selecting research problems.
To install it, open Cowork, go to the plugin directory, and install Bio Research. In Claude Code, add the anthropics/knowledge-work-plugins marketplace and install bio-research from it. Run /start to see which tools are connected.
Where Bio Research stops
The plugin is scoped to preclinical work: genomics, target identification and literature search. Full-text reading depends on your institution's journal subscriptions, and connectors such as Benchling need their own accounts. Treat it as a strong default for discovery-stage biology rather than a clinical or regulatory toolkit.
K-Dense Scientific Agent Skills: the broadest Claude plugin for computational biology
K-Dense's Scientific Agent Skills is the largest open collection of scientific skills for Claude. As of July 2026, the repository lists more than 160 skills and access to more than 78 scientific databases. It covers cancer genomics, variant-effect prediction, RNA velocity, molecular dynamics, PK/PD modelling, cheminformatics and biomedical literature retrieval. The repository is packaged as a portable plugin, so Claude Code and Cowork can load the whole collection at once. You can also install individual skills with gh skill install K-Dense-AI/scientific-agent-skills.
Its strength is breadth for people who write code. A bioinformatician can ask Claude to run Scanpy clustering, compute RDKit descriptors and query a variant database in one session, without wiring each library by hand. K-Dense also publishes a smaller plugin, Claude Scientific Writer, with a pinned subset of 26 skills for manuscript drafting.
Install third-party science skills like you would install code
Skills can run scripts and install Python packages on your machine. K-Dense itself asks users to read its security guidance before installing skills in a lab environment. Before installing, pin a version, read the skill files, and check your institution's data policy before running them on unpublished or patient-derived data.
Claude Science: a dedicated app for biomedical research, not a plugin
Some researchers searching for "Claude plugins for biology" actually need Claude Science. Anthropic launched it on 30 June 2026 as a beta desktop app for macOS and Linux on Pro, Max, Team and Enterprise plans. It is a separate research workspace rather than an add-on to Cowork.
The app comes with more than 60 optional scientific databases and renders protein structures, genome browser tracks and chemical structures natively. It records how each figure and number was produced, and a reviewer agent flags citations and numbers it cannot trace. In August 2026, Anthropic opened 10,000 free and discounted seats for scientists through a Claude Team plan for research labs.
Choose Claude Science if you want one environment for analysis and drafting. Choose plugins if you already work in Cowork or Claude Code and want to add specific capabilities.
Verifying biomedical citations: the layer life-science plugins leave out
Without retrieval, language models invent most biomedical citations. In the OpenScholar benchmark, 78–98% of paper titles cited by non-retrieval models did not exist, and biomedicine was the worst domain: 94.8% for GPT-4o and 96.6% for Llama 3.1 70B (PMID 41639446). Earlier studies found 55% of GPT-3.5 citations and 18% of GPT-4 citations fabricated (PMID 37679503), and 47% fabricated in ChatGPT-generated medical content (PMID 37337480).
Retrieval largely solves the existence problem. OpenScholar-8B, which retrieves from a 45-million-paper store and checks its own citations, produced no hallucinated titles (PMID 41639446). Almanac, which retrieves from curated clinical sources including PubMed, gave correct citations for 91% of 314 clinical questions across nine specialties (PMID 38343631). A PubMed connector gives Claude the same basic protection: it cites records that exist.
Retrieval does not solve the attribution problem. In a Stanford red-teaming study of 1,504 responses, 20.1% were inappropriate, and requests for citations were among the prompts most likely to produce hallucinations (PMID 40055532). In one case, GPT-4 cited three pain-management papers whose authors, titles, journals and years were all correct, yet none supported the recommendation it made. None of the plugins above runs a claim-by-claim check against the source text.
That check is what the BioSkepsis connector adds. BioSkepsis is not a plugin: it is an MCP connector that works in Claude.ai, Claude Desktop and Cowork, including inside a Cowork task alongside Bio Research. A research run does the following:
- Searches a corpus of more than 40 million biomedical papers and reads full text where it is available.
- Screens out retracted and questioned papers.
- Checks every cited claim against the paper it is attributed to.
- Returns a brief with a Trust Index and a research notebook, which lists any failed citations separately.
The trade-off is time: a run takes several minutes, against seconds for a PubMed lookup. Setup is covered in how to connect Claude to BioSkepsis, and a worked run is in how to use the BioSkepsis connector.
What verification caught in the research run behind this post
The draft brief stated that retrieval reduced hallucinated titles to zero for every model tested. Verification against the OpenScholar paper found that the zero applied only to OpenScholar-8B, so the claim was struck. Of 26 cited supports in the run, 19 were verified against full text, 3 were struck and 4 rested on abstracts only. Every figure in this section comes from the verified supports.
Using BioSkepsis inside a plugin workflow
Connect BioSkepsis once under Customize → Connectors, either from the connector directory or as a custom connector with the server URL https://app.bioskepsis.ai/api/mcp. Cowork tasks can then call it next to any installed plugin. Plugins are file-based, so an organisation can also add the same URL to the .mcp.json of its own customised copy of Bio Research. Team members then get verified literature review as part of the same install.
Comparing Claude plugins for life-science research by task
| Task | Bio Research plugin | K-Dense Scientific Agent Skills | Claude Science app | BioSkepsis connector |
|---|---|---|---|---|
| Type | Plugin (Anthropic-verified) | Plugin / skill collection (open source) | Standalone beta app | MCP connector |
| Runs in | Cowork, Claude Code | Claude Code, Cowork | Claude Science (macOS, Linux) | Claude.ai, Desktop, Cowork |
| Literature search | PubMed, bioRxiv, Wiley | Via literature-retrieval skills | Built-in databases | 40M+ papers, full text where available |
| Omics analysis | scRNA-seq QC, scvi-tools, nf-core | Broadest: Scanpy, RNA velocity, variant effects and more | Genomics, single-cell, proteomics | No |
| Target and compound data | Open Targets, ChEMBL | Cheminformatics and drug-target skills | Cheminformatics, structural biology | Literature evidence on targets only |
| Claim-level citation check | No | No | Reviewer agent flags untraceable citations | Yes: each claim checked against its source; failures struck and listed |
| Retraction screening | No | No | Not stated | Yes |
| Main limitation | Preclinical scope; full text depends on subscriptions | Third-party code to vet; command-line setup | Beta; desktop only | 3 to 8 minutes per run; one monthly brief each |
These options complement rather than compete with each other. A preclinical team typically runs Bio Research for data and databases, adds K-Dense for heavier computational work, and routes any claim headed for a manuscript or grant through a verified literature step.
Which Claude plugin setup fits which life-science researcher
Bio Research Wet-lab biologists and translational scientists
Install one plugin and get PubMed, bioRxiv, Open Targets, ChEMBL and ClinicalTrials.gov in the same Cowork task, plus scRNA-seq QC for the data coming off your own experiments.
K-Dense Bioinformaticians and computational biologists
Load the scientific skills collection in Claude Code when the work is code-heavy: clustering, variant annotation, molecular descriptors and pipeline debugging across many libraries.
BioSkepsis PhD students, medical writers and grant authors
Add the BioSkepsis connector to whichever setup you use. Any claim headed for a thesis chapter, manuscript or grant then arrives with each citation checked against its source and retracted papers screened out.
Claude Science Labs that want a dedicated research workspace
If your group is starting fresh and works on macOS or Linux, a single app with built-in databases and traceable artifacts may suit you better than assembling plugins.
Claude plugins for life-science research: frequently asked questions
What is the best Claude plugin for life-science research?
For most wet-lab and translational researchers, Anthropic's Bio Research plugin is the right first install. It bundles 10 MCP integrations, including PubMed, bioRxiv, ClinicalTrials.gov, ChEMBL, Open Targets and Benchling, with 6 analysis skills. Computational biologists working in Claude Code get more breadth from K-Dense's Scientific Agent Skills, which package more than 160 skills as one plugin.
Are Claude plugins free for researchers?
The plugins covered here are open source and free to install. Using them requires a Claude plan that includes Cowork or Claude Code. Some connectors inside a plugin also need their own account or institutional access, such as Benchling, Medidata or full-text journal subscriptions.
What is the difference between a Claude plugin and a Claude connector?
A connector links Claude to one external service through the Model Context Protocol and gives Claude that service's tools. A plugin is a package for Cowork and Claude Code that can bundle several connectors with skills, slash commands and subagents. Connectors also work in Claude.ai chat; plugins run in Cowork and Claude Code.
Is BioSkepsis a Claude plugin?
No. BioSkepsis is a Claude connector built on the Model Context Protocol. It works in Claude.ai, Claude Desktop and Cowork alongside any installed plugin. An organisation can also add the BioSkepsis server URL to its own customised plugin.
Does the PubMed connector stop Claude from inventing citations?
It largely removes the risk of citing papers that do not exist: without retrieval, 78–98% of titles cited by language models were invented in one Nature benchmark. It does not check whether a retrieved paper supports the specific sentence it is attached to, whether the paper has been retracted, or how strong the evidence is. Those checks need a separate verification step.
Should I use Claude Science instead of plugins?
Claude Science is a separate beta desktop app for macOS and Linux, launched in June 2026. It has more than 60 optional scientific databases, and its artifacts are traced back to the code that produced them. It suits researchers who want one workspace for analysis and drafting. Plugins suit researchers who already work in Cowork or Claude Code and want to add specific capabilities.
Are third-party Claude plugins safe to install in a lab environment?
Treat them like any other code you run. Skills can execute scripts and install packages on your machine, so read the plugin's files before installing, prefer Anthropic-verified or widely reviewed repositories, and pin versions. Check your institution's data policy before pointing Claude at unpublished or patient-derived data.
Add verified biomedical literature to your Claude setup
Keep the plugins you use for data and databases, and connect BioSkepsis for the claims that end up in your manuscript. Every brief arrives with citations checked against their sources, retracted papers screened out, and a Trust Index you can inspect.
Start freeSources for the Claude life-science plugin comparison
- Walters WH, Wilder EI. Fabrication and errors in the bibliographic citations generated by ChatGPT. Sci Rep. 2023;13:14045. PMID 37679503. Cited for: 636 citations across 84 generated reviews; 55% fabricated with GPT-3.5 and 18% with GPT-4; substantive errors in 43% and 24% of real citations.
- Bhattacharyya M, Miller VM, Bhattacharyya D, Miller LE. High rates of fabricated and inaccurate references in ChatGPT-generated medical content. Cureus. 2023;15(5):e39238. PMID 37337480. Cited for: 115 references, 47% fabricated, 46% real but inaccurate, 7% real and accurate; incorrect PMID in 93% of papers.
- Asai A, et al. Synthesizing scientific literature with retrieval-augmented language models. Nature. 2026;650(8103):857-863. PMID 41639446. Cited for: 78–98% non-existent titles from non-retrieval models, 94.8% (GPT-4o) and 96.6% (Llama 3.1 70B) in biomedicine, and no hallucinated titles from OpenScholar-8B.
- Zakka C, et al. Almanac: retrieval-augmented language models for clinical medicine. NEJM AI. 2024;1(2). PMID 38343631. Cited for: correct citations for 91% of 314 clinical questions across nine specialties.
- Chang CT, et al. Red teaming ChatGPT in medicine to yield real-world insights on model behavior. npj Digit Med. 2025;8:149. PMID 40055532. Cited for: 1,504 responses, 20.1% inappropriate, citation requests as a hallucination-prone prompt type, and real citations that did not support the claim.
- Anthropic. Claude for Life Sciences (October 2025). anthropic.com. Cited for: launch connectors Benchling, BioRender, PubMed, Wiley Scholar Gateway, Synapse.org and 10x Genomics.
- Anthropic. Advancing Claude in healthcare and the life sciences (January 2026). anthropic.com. Cited for: Medidata, ClinicalTrials.gov, ToolUniverse, bioRxiv/medRxiv, Open Targets, ChEMBL and Owkin connectors.
- Anthropic. Bio Research plugin. claude.com/plugins/bio-research and github.com/anthropics/knowledge-work-plugins. Cited for: 10 MCP integrations, 6 analysis skills, skill scope and install route.
- K-Dense. Scientific Agent Skills. github.com/K-Dense-AI/scientific-agent-skills. Cited for: more than 160 skills, 78+ databases, portable plugin packaging and install command.
- K-Dense. Claude Scientific Writer. github.com/K-Dense-AI/claude-scientific-writer. Cited for: 26-skill pinned subset and security guidance for lab environments.
- Anthropic. Claude Science, an AI workbench for scientists (30 June 2026). anthropic.com. Cited for: beta availability, platforms and plans, traceable artifacts.
- Anthropic. Expanding our support for scientists (2026). anthropic.com. Cited for: 10,000 free and discounted seats through the Claude Team plan for scientists.
- BioSkepsis. Claude connector documentation. bioskepsis.ai/docs/claude-connector. Cited for: MCP server URL, corpus size, run time and tools.
- Related reading on the BioSkepsis blog: How to connect Claude to BioSkepsis, How to use the BioSkepsis connector in Claude and 10 best AI tools for life-science literature review.