I want to share Pipette.bio, a pay-as-you-go bioinformatics platform we have been building. You describe an analysis in plain English, the agent proposes a plan, you approve it, and it writes and runs the actual code (bash / Python / R) on cloud workers, then hands back figures, tables, and a reproducible report. No coding required, but every command, version, and parameter is captured so the work stays auditable.

Link: https://pipette.bio Contact: support@pipette.bio

What it is

Pipette pairs an LLM (Claude) with standard, industry bioinformatics tooling running in real Conda environments. It is not a chatbot that writes code for you to run elsewhere: it executes the code itself on managed compute, inspects the outputs, handles errors, and iterates until the analysis is done.

How a run works (plan-approval gate)

Nothing executes until you approve a plan. This is the core safety feature and the thing that stops the agent from burning compute on the wrong analysis.

1. Setup - your input files are staged from cloud storage into a fresh workspace sized for the job.
2. Planner - a dedicated planning agent reads your request, the conversation, and your files, then proposes a structured plan: which skill, which steps, which parameters, and which assumptions it is making. Each parameter is traced to its source (a phrase you typed, an earlier turn, a tool default, or an inference). You Approve or Reject / revise with feedback.
3. Agentic execution - after approval the executor is locked to the plan and runs it step by step.
4. Methodology review - an independent reviewer agent checks statistical correctness (tests, normalization, multiple-testing correction, QC filtering) and plan compliance, and triggers a remediation loop if something is off.
5. Provenance - tool versions, parameters, input checksums, and command transcripts are recorded.
6. Literature and hypotheses - a literature-aware agent reads the outputs, retrieves relevant PubMed papers, and returns narrative context plus suggested follow-ups.
7. Finalize - results are archived to persistent storage and a completion summary is posted.

Key features

• Natural-language interface. Describe the goal, the input files, and key parameters; the agent designs the workflow.
• Plan-before-execute approval gate. You see and approve the exact plan (steps, parameters, assumptions) before any compute is spent.
• Independent methodology review with auto-remediation. A second agent audits the analysis and re-runs affected steps if it finds problems.
• Over 100 specialized analysis skills spanning bulk RNA-seq, single-cell, spatial omics, variant calling, cancer genomics, GWAS and population genetics, microbiome and metagenomics, epigenomics, methylation, proteomics, phylogenetics, comparative genomics, drug design, CRISPR screens, and synthetic biology.
• 150+ underlying tools in versioned Conda environments (STAR, HISAT2, Salmon, BWA, GATK4, DESeq2, edgeR, limma, Seurat, Scanpy, kraken2, QIIME2, PLINK/PLINK2, AutoDock Vina, GROMACS, and many more).
• 15 external knowledge databases through an automatic query router: PubMed, ClinicalTrials.gov, ClinVar, gnomAD, NCBI Nucleotide, GEO, UniProt, dbSNP, AlphaFold, cBioPortal, UCSC, Ensembl, Open Targets, OpenFDA, and the GWAS Catalog.
• Purpose-built compute environments. Separate worker classes for general NGS, microbiome, and single-cell workloads, plus larger queues for RNA-seq alignment and big single-cell jobs.
• Direct data loading from public URLs. Dropbox, Google Drive, Box, OneDrive, Zenodo, Figshare, OSF, Hugging Face, GitHub raw, plain HTTP/HTTPS, FTP/SFTP, and S3 presigned links. Or upload directly, or pull straight from SRA/GEO.
• Automatic, reproducible reports. AI-generated methods, results, and figures, with full provenance (versions, parameters, input hashes, command transcripts).
• Persistent cloud storage and history. 500 GB per account, outputs organized by Task ID and step, full session transcripts you can resume or export as PDF.
• Multi-session and background tasks. Run several analyses at once; jobs in other sessions keep running and show live progress.

Domain coverage (selected)

• Bulk RNA-seq: QC, alignment (STAR/HISAT2) or alignment-free quant (Salmon/kallisto), DESeq2 / edgeR / limma DE, GO/KEGG/GSEA enrichment, WGCNA, DTU/DEXSeq, splicing (rMATS), fusion detection.
• Single-cell and spatial: Scanpy and Seurat pipelines, doublet detection, batch correction (Harmony/BBKNN), trajectory inference (Monocle3/Slingshot/CellRank), RNA velocity, cell-cell communication (CellChat/CellPhoneDB), SingleR/CellTypist annotation, scATAC (Signac/ArchR), Visium / Visium HD / Xenium / CosMx / Slide-seq / MERFISH, deconvolution (cell2location/RCTD/Tangram).
• Variants and cancer: GATK / FreeBayes germline, Mutect2 somatic, SnpEff annotation, CNVkit, Arriba fusions, MSIsensor-pro, maftools cohort summaries, structural variants (Delly/SURVIVOR), sample-swap and contamination QC (somalier/VerifyBamID2).
• Population genetics and breeding: PLINK/PLINK2 GWAS, ADMIXTURE, GCTA heritability, SuSiE fine-mapping, Beagle imputation, TASSEL, rrBLUP/BGLR genomic prediction, R/qtl mapping, HLA typing (T1K).
• Microbiome and metagenomics: kraken2/bracken, MetaPhlAn/HUMAnN, QIIME2 amplicon, assembly (SPAdes/MEGAHIT/Flye), binning (MetaBAT2/MaxBin2/DAS_Tool), dRep, GTDB-Tk, pangenomes (Roary), virome (VirSorter/CheckV/Pharokka).
• Epigenomics and methylation: MACS3 peak calling, DiffBind, ChIPseeker, deepTools, ATAC-seq, CUT&RUN / CUT&Tag, Bismark + methylKit/DMRseq/DSS, methylation age clocks, MEME-suite motif discovery.
• Drug design and cheminformatics: AutoDock Vina/SMINA docking and virtual screening, RDKit, fpocket, PLIP/ProLIF interaction analysis, ADMET-AI, GROMACS/OpenMM MD, ChEMBL/PubChem/RCSB queries.
• Proteomics, survival, phylogenetics, comparative genomics, CRISPR screens, synthetic biology and more.

Who it is for

Wet-lab biologists and clinicians who want analyses without writing code, and bioinformaticians who want to offload routine pipelines while keeping full provenance and the ability to revise the plan.

Happy to answer questions here. Feedback on missing tools or workflows is very welcome.