This is a space for depositing various helpful artifacts (skills, commands, etc.) for various agentic things like Claude Code or Gemini Code Agent. The idea is that by iterating over this we will create a robust set of community curated "skills" that would ultimately create guardrails preventing LLMs from doing crazy 💩 and ensuring best practices. Should probably be called Intergalactic LLM-taming commision (ILT).
This was not written by Claude.
this is complementary to https://github.com/galaxyproject/galaxy-mcp
Skills are structured instructions that teach an agent how to actually use Galaxy -- run analyses on a real instance, keep the work on-graph and reproducible, and extend Galaxy when the tool you need isn't there yet. They provide:
- Best practices - The right way to run, structure, and hand off an analysis
- Patterns - Common solutions to common problems
- Examples - Concrete references to learn from
- Guardrails - Preventing common mistakes
The analysis runs on Galaxy, not on your laptop. The agent drives a real Galaxy server over MCP, so compute happens on Galaxy's infrastructure and every step lands in a history with full provenance -- a durable, shareable record rather than a folder of loose files. That includes writing new tools when one is missing: User-Defined Tools let the agent add a custom step to an analysis without leaving Galaxy or waiting on an admin.
| Repository | Purpose | Use For |
|---|---|---|
| skills (this repo) | Knowledge - How to use Galaxy well | Running analyses, structuring collections, authoring User-Defined Tools, reproducible handoff |
| galaxy-mcp | Capabilities - Interact with Galaxy programmatically | The live connection the skills drive |
Use both together -- the skills describe how to drive the connection galaxy-mcp provides.
Clone into the skills directory your agent reads:
cd ~/.claude/skills
git clone https://github.com/galaxyproject/galaxy-skills galaxyPackaged installs -- a marketplace entry that also wires up the MCP server -- are assembled in a
separate distribution repository, which vendors the skills/ tree from here. This repository stays
the source of the skills themselves.
# Install openskills
npm i -g openskills
# Install Galaxy skills
openskills install galaxyproject/galaxy-skills
# Load a specific skill when needed
openskills read galaxy-mcp-referenceClone this repo into your workspace and reference skills in your prompts:
git clone https://github.com/galaxyproject/galaxy-skillsThe LLM can read skill files directly from the workspace.
The skills drive a live server through galaxy-mcp,
which runs under uvx and needs an instance and a key:
export GALAXY_URL=https://usegalaxy.org # this is the default if unset
export GALAXY_API_KEY=your_api_key # User -> Preferences -> Manage API KeyA .env file in your working directory works too, and credentials can also be supplied at runtime
with connect(url, api_key). The server exposes 45 tools; GALAXY_MCP_EXCLUDE_TAGS and
GALAXY_MCP_INCLUDE_TAGS trim what sits in context. skills/galaxy-integration/ covers the setup
and the alternatives (JupyterLite, BioBlend) in more detail.
skills/ holds the skills for using Galaxy -- running an analysis on a real instance, keeping it
structured and reproducible, and extending Galaxy when a tool you need is missing. This is the tree
a packaged install ships.
galaxy-integration -- Connect an agent to a Galaxy instance and choose how to drive it (MCP, JupyterLite notebooks, or BioBlend). Start here.
galaxy-mcp-reference -- The Galaxy MCP tool surface: histories, datasets, tools, workflows, invocations, and the pitfalls that bite first.
collection-manipulation -- Transform Galaxy dataset collections reproducibly with native tools: filter, sort, relabel, merge, flatten, nest, and the Apply Rules DSL.
udt-authoring -- Author User-Defined Tools: a class: GalaxyUserTool YAML definition wrapping a
container and command into a tool a non-admin user creates and runs. This is how an agent adds a
custom analysis step without leaving Galaxy or waiting on an admin.
workflow-reports -- Write workflow report templates for the Workflow Editor's Report tab.
reproduciblify -- Re-execute a messy, ad-hoc history as a clean, fully on-graph, collection-structured analysis that extracts into a reusable workflow.
dev-skills/ holds skills for building Galaxy rather than using it -- tool-dev (authoring tool
XML wrappers for tools-iuc), nf-to-galaxy (converting Nextflow processes and pipelines),
update-usegalaxy-tool (ToolShed revisions in usegalaxy-tools), trackhubs (publishing UCSC Track
Hubs, which is mostly UCSC tooling outside Galaxy), and hub-news-posts (galaxyproject.org news).
They sit outside skills/ so that installing the analysis set doesn't hand an analyst a pile of
tool-development guidance, but they remain here to read, to link to, and to install by hand.
galaxy-skills/
├── AGENTS.md # Cross-agent routing instructions
├── CONTRIBUTING.md # How to add new skills
│
├── skills/ # <- using Galaxy; this is what gets packaged
│ ├── galaxy-integration/ # Connect and choose how to drive Galaxy
│ │ └── jupyterlite/ # JupyterLite notebooks (gxy package)
│ ├── galaxy-mcp-reference/ # Galaxy MCP tool surface
│ ├── collection-manipulation/ # Collection transformations
│ ├── udt-authoring/ # User-Defined Tools
│ ├── workflow-reports/ # Workflow report templates
│ └── reproduciblify/ # Ad-hoc history -> reusable workflow
│
└── dev-skills/ # <- building Galaxy; not registered by any harness
├── tool-dev/ # Authoring Galaxy tool XML wrappers
├── nf-to-galaxy/ # Nextflow -> Galaxy conversion
├── trackhubs/ # UCSC Track Hub publishing
├── update-usegalaxy-tool/ # ToolShed revisions in usegalaxy-tools
└── hub-news-posts/ # galaxyproject.org news posts
Every major agent harness -- Claude Code, Codex, and Antigravity -- discovers skills at skills/
by convention, and none of them scans dev-skills/. That is what keeps an install focused on
analysis without deleting anyone's work.
We welcome contributions! See CONTRIBUTING.md for:
- How to create a new skill
- Skill structure and format
- Testing and validation
- Submission guidelines
User: "Upload these FASTQ files and run FastQC on all of them"
AI: [Loads galaxy-integration to connect, then galaxy-mcp-reference]
[Creates a history, uploads the reads]
[Builds a list collection so one job covers every sample]
[Runs FastQC mapped over the collection, reports when the jobs finish]
User: "There's no Galaxy tool for this container -- can I still run it?"
AI: [Loads udt-authoring skill]
[Writes a `class: GalaxyUserTool` YAML wrapping the container]
[Creates it with create_user_tool and runs it -- no admin required]
User: "Turn this history into something I can rerun on new samples"
AI: [Loads reproduciblify skill]
[Re-executes the analysis on-graph with collection structure]
[Authors a notebook that extracts into a sample-agnostic workflow]
- galaxy-mcp - MCP server for Galaxy interaction
- planemo - Galaxy tool development toolkit
- Galaxy - Main Galaxy platform
- IWC - Intergalactic Workflow Commission