SciAgent-Skills MCP — Use Cases, Install & Live Demo

Why use it

Key features

197 specialized life-science skills
Coverage: RNA-seq, single-cell (scRNA), drug discovery, proteomics, variant calling
BixBench accuracy 92.0% — higher than generic agents
Canonical tool invocation (samtools, bcftools, salmon, scanpy, Seurat, etc.)
Reproducibility patterns (conda/mamba envs, config files)
Powers OmicsHorizon platform

Live Demo

What it looks like in practice

sciagent-skill.replay ▶ ready

0/0

Install

Pick your client

~/Library/Application Support/Claude/claude_desktop_config.json · Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "sciagent-skill": {
      "command": "git",
      "args": [
        "clone",
        "https://github.com/jaechang-hits/SciAgent-Skills",
        "~/.claude/skills/SciAgent-Skills"
      ],
      "_inferred": true
    }
  }
}

Open Claude Desktop → Settings → Developer → Edit Config. Restart after saving.

~/.cursor/mcp.json · .cursor/mcp.json

{
  "mcpServers": {
    "sciagent-skill": {
      "command": "git",
      "args": [
        "clone",
        "https://github.com/jaechang-hits/SciAgent-Skills",
        "~/.claude/skills/SciAgent-Skills"
      ],
      "_inferred": true
    }
  }
}

Cursor uses the same mcpServers schema as Claude Desktop. Project config wins over global.

VS Code → Cline → MCP Servers → Edit

{
  "mcpServers": {
    "sciagent-skill": {
      "command": "git",
      "args": [
        "clone",
        "https://github.com/jaechang-hits/SciAgent-Skills",
        "~/.claude/skills/SciAgent-Skills"
      ],
      "_inferred": true
    }
  }
}

Click the MCP Servers icon in the Cline sidebar, then "Edit Configuration".

~/.codeium/windsurf/mcp_config.json

{
  "mcpServers": {
    "sciagent-skill": {
      "command": "git",
      "args": [
        "clone",
        "https://github.com/jaechang-hits/SciAgent-Skills",
        "~/.claude/skills/SciAgent-Skills"
      ],
      "_inferred": true
    }
  }
}

Same shape as Claude Desktop. Restart Windsurf to pick up changes.

~/.continue/config.json

{
  "mcpServers": [
    {
      "name": "sciagent-skill",
      "command": "git",
      "args": [
        "clone",
        "https://github.com/jaechang-hits/SciAgent-Skills",
        "~/.claude/skills/SciAgent-Skills"
      ]
    }
  ]
}

Continue uses an array of server objects rather than a map.

~/.config/zed/settings.json

{
  "context_servers": {
    "sciagent-skill": {
      "command": {
        "path": "git",
        "args": [
          "clone",
          "https://github.com/jaechang-hits/SciAgent-Skills",
          "~/.claude/skills/SciAgent-Skills"
        ]
      }
    }
  }
}

Add to context_servers. Zed hot-reloads on save.

claude mcp add sciagent-skill -- git clone https://github.com/jaechang-hits/SciAgent-Skills ~/.claude/skills/SciAgent-Skills

One-liner. Verify with claude mcp list. Remove with claude mcp remove.

Use Cases

Real-world ways to use SciAgent-Skills

How to run a bulk RNA-seq differential expression analysis

👤 Biologists and bioinformaticians with FASTQ files and a need for DE results ⏱ ~120 min advanced

When to use: You have paired-end FASTQ, a reference, and a condition table.

Prerequisites

Skill installed — git clone https://github.com/jaechang-hits/SciAgent-Skills ~/.claude/skills/sciagent-skills
Conda/mamba + compute — Use a Linux server or HPC; macOS fine for small datasets

Flow

Plan the pipeline

Plan a bulk RNA-seq DE workflow for 2 conditions x 3 reps with salmon + tximport + DESeq2. Output expected files per step.✓ Copied

→ Step list with tools and intermediate outputs
Run quantification

Run salmon quant for each sample; produce a script.✓ Copied

→ Bash script with salmon invocations
DE analysis

Load quants via tximport, run DESeq2, produce MA + volcano plots and a top-50 gene table.✓ Copied

→ R script + output files

Outcome: A DE table + plots you can hand to a PI or collaborator.

Pitfalls

Mis-matched reference versus annotation GTF — Check release numbers explicitly; salmon and DESeq2 can silently run on mismatched IDs
Low-count filtering removes biological signal — Use independent filtering or raise threshold gradually; don't blanket filter

Single-cell RNA-seq QC and clustering baseline

👤 Researchers starting scRNA analysis from 10x output ⏱ ~90 min advanced

When to use: You have a cellranger output directory and need a first-pass Seurat or Scanpy analysis.

Flow

Load + QC

Load 10x data from cellranger output, do QC (pct.mt, counts, features), filter, normalize.✓ Copied

→ QC plots + filtered object
Cluster

Run PCA, find neighbors, cluster at resolution 0.5, and UMAP.✓ Copied

→ Cluster labels + UMAP plot
Marker genes

Find cluster markers, produce a top-5-per-cluster heatmap.✓ Copied

→ Marker table + heatmap

Outcome: An annotated UMAP and marker table ready for biological interpretation.

Pitfalls

Over-clustering at the default resolution — Try a sweep 0.2-1.0 and pick based on silhouette + biological plausibility

Germline variant calling from WGS

👤 Genomics teams calling SNPs and indels ⏱ ~180 min advanced

When to use: You have aligned BAMs and need a VCF using best-practices (GATK).

Flow

Plan pipeline

Plan GATK best-practices germline pipeline from aligned BAMs: BQSR, HaplotypeCaller, joint genotyping with GenomicsDBImport.✓ Copied

→ Pipeline outline with expected runtimes
Generate scripts

Produce Snakemake rules for each step.✓ Copied

→ Snakefile with rules and config

Outcome: A reproducible Snakemake pipeline ready to run on your cluster.

Pitfalls

Skipping BQSR on small cohorts — Still do it — GATK's downstream filters assume recalibrated quality scores

Triage small-molecule hits from a screen

👤 Med-chem groups triaging HTS results ⏱ ~60 min advanced

When to use: You have a hit list and want to rank by drug-likeness + novelty.

Flow

Filter by Lipinski + PAINS

Compute Lipinski and PAINS flags on this SMILES list, output a filtered table.✓ Copied

→ RDKit-based script + filtered CSV
Similarity to known drugs

For remaining hits, compute Tanimoto similarity to ChEMBL approved drugs; flag >0.85 as known-scaffold.✓ Copied

→ Similarity table with flags

Outcome: A triaged hit list prioritized for follow-up.

Combinations

Pair with other MCPs for X10 leverage

sciagent-skill + filesystem

Read local FASTQ/BAM files and have SciAgent plan the analysis pipeline on actual data

List the FASTQ files in ~/data/rnaseq/, then use SciAgent to plan a salmon + DESeq2 pipeline for these samples.✓ Copied

sciagent-skill + github

Search nf-core or Bioconductor repos for reference implementations, then adapt with SciAgent

Search nf-core/rnaseq on GitHub for how they handle trimming, then use SciAgent to adapt that step for my pipeline.✓ Copied

Tools

What this MCP exposes

Tool	Inputs	When to call	Cost
Bulk RNA-seq workflow	FASTQ + design	Standard DE analysis	compute
scRNA-seq workflow	10x output	Single-cell baseline	compute
Variant calling	BAMs	WGS/WES cohorts	compute
Proteomics analysis	MS data	MS-based proteomics	compute
Drug discovery triage	SMILES list	Hit triage	compute

Cost & Limits

What this costs to run

API quota: None at skill level
Tokens per call: 10-50k per pipeline design
Monetary: Free skills; compute costs depend on dataset size
Tip: Plan the pipeline as a script first, then run; don't keep Claude in the loop during long compute.

Security

Permissions, secrets, blast radius

Credential storage: No credentials

Data egress: Pipeline designs and snippets go to Claude. Actual patient data should stay on HIPAA-compliant compute.

Do not paste patient identifiers or raw sequence data that could be re-identified into prompts.
Always review generated variant-calling parameters — wrong caller settings can produce misleading results.

Troubleshooting

Common errors and fixes

tximport fails on missing transcript-to-gene mapping

Regenerate tx2gene from the same GTF you used for the salmon index; mismatch is the usual cause.

Seurat object too large for available RAM

Subsample or switch to on-disk storage with BPCells or DelayedArray-backed Seurat

Conda environment conflicts during pipeline setup

Use mamba instead of conda for faster resolution; pin exact versions in environment.yml rather than floating constraints

Verify: mamba env create -f environment.yml --dry-run

Generated script references wrong genome build (hg19 vs hg38)

Explicitly state the genome build in your first prompt; SciAgent defaults may not match your data

Alternatives

SciAgent-Skills vs others

Alternative	When to use it instead	Tradeoff
Galaxy / nf-core	You want audited, community pipelines rather than LLM-generated scripts	Less conversational; slower to customize

More

Resources

📖 Read the official README on GitHub

🐙 Browse open issues

🔍 Browse all 400+ MCP servers and Skills