Galaxy agentic plugins

Plugins and setup guides that connect AI coding agents to Galaxy. One install gives an agent:

1. A Galaxy connection through the galaxy-mcp MCP server, authenticated with your Galaxy API key.
2. Curated Galaxy skills from galaxyproject/galaxy-skills, split the way upstream splits them: skills for using Galaxy (MCP tool surface, collections, user-defined tools, workflow reports, reproducibility) and, separately, skills for building it (tool wrappers, Nextflow conversion, ToolShed revisions, track hubs, hub posts).
3. Galaxy Workflow Foundry skills from galaxyproject/foundry (papers, Nextflow and CWL to validated Galaxy workflows).

Pick your harness

Harness	Guide	Install in one line
Claude Code	docs/claude-code.md	/plugin marketplace add galaxyproject/agentic-plugins then /plugin install galaxy-mcp@galaxyproject
Claude Desktop	docs/claude-desktop.md	Open galaxy-mcp.mcpb from the latest release
Codex CLI	docs/codex.md	codex plugin marketplace add galaxyproject/agentic-plugins then codex plugin add galaxy-mcp@galaxyproject
Cursor	docs/cursor.md	Customize > From GitHub Repository > galaxyproject/agentic-plugins, or the one-click MCP link in the guide
Antigravity (agy)	docs/antigravity.md	git clone this repo, then agy plugin install <clone>/plugins/galaxy-mcp
Pi	docs/pi.md	pi install npm:pi-mcp-adapter then pi install git:github.com/galaxyproject/agentic-plugins

Every guide starts with getting a Galaxy API key.

What is in the box

Four plugins live under plugins/. Each directory is simultaneously a Claude Code plugin (.claude-plugin/plugin.json), a Codex plugin (.codex-plugin/plugin.json), a Cursor plugin (.cursor-plugin/plugin.json) and an Antigravity plugin (plugin.json); the root package.json exposes the same directories as a Pi package. bundles/claude-desktop/ is a separate one-click installer for Claude Desktop (MCP server only).

Plugin	Contents	Origin
galaxy-mcp	MCP server config for uvx galaxy-mcp (mcp/claude.json, mcp/pi.json, mcp.json for Cursor, mcp_config.json for Antigravity, inline in the Codex manifest) and the galaxy-connect skill: set up, verify and troubleshoot the connection	this repo
galaxy-skills	7 skills for using Galaxy: galaxy-integration (+jupyterlite), galaxy-mcp-reference, collection-manipulation, udt-authoring, workflow-reports, reproduciblify	mirror of galaxyproject/galaxy-skills skills/
galaxy-dev-skills	9 skills for building Galaxy: tool-dev (+tool-selection-diagram), nf-to-galaxy (+3 sub-skills), update-usegalaxy-tool, trackhubs, hub-news-posts	mirror of galaxyproject/galaxy-skills dev-skills/
foundry-skills	59 cast skills, e.g. pipeline-nextflow-to-galaxy, discover-shed-tool, validate-galaxy-workflow, author-galaxy-tool-wrapper	mirror of galaxyproject/foundry casts/claude/skills

Install only the plugins you need. Every skill plugin adds its descriptions to the agent's context on every session: galaxy-skills is the one an analyst wants, galaxy-dev-skills is for people who write tool wrappers, and foundry-skills is only worth it if you build Galaxy workflows.

How credentials reach the MCP server

galaxy-mcp reads GALAXY_URL and GALAXY_API_KEY from its environment or a project .env file. Harnesses differ in how the environment gets there, which is why the plugin ships one MCP file per harness:

Harness	Mechanism
Claude Code	Prompted at install time (userConfig), stored in the keychain, injected via ${user_config.*}
Codex CLI	Only allow-listed variables are forwarded; the Codex manifest lists the two (env_vars)
Cursor	${env:GALAXY_URL} / ${env:GALAXY_API_KEY} interpolated from the environment in mcp.json
Antigravity	Shell environment inherited as-is; ${VAR} in mcp_config.json is not expanded
Pi	pi-mcp-adapter inherits the shell environment
Claude Desktop	Prompted at install time (bundle user_config), stored in the keychain

Repository layout

.claude-plugin/marketplace.json   Claude Code marketplace "galaxyproject"
.agents/plugins/marketplace.json  Codex marketplace "galaxyproject"
.cursor-plugin/marketplace.json   Cursor marketplace "galaxyproject"
package.json                      Pi package manifest (pi.skills, pi.mcp)
plugins/
  galaxy-mcp/                     MCP config + galaxy-connect skill (hand-written)
  galaxy-skills/                  vendored copy of galaxy-skills skills/ (using Galaxy)
  galaxy-dev-skills/              vendored copy of galaxy-skills dev-skills/ (building Galaxy)
  foundry-skills/                 vendored copy of galaxyproject/foundry casts/claude/skills
bundles/claude-desktop/           Claude Desktop .mcpb source (uv runtime, no bundled code)
docs/                             per-harness guides (source for the Galaxy Hub pages)
scripts/
  sync-skills.sh                  re-vendor upstream skills, regenerate manifests, validate
  gen-manifests.py                writes generated manifest fields (nested skill lists, versions)
  validate.py                     structural checks (run in CI)
  mcp_smoke.py                    starts galaxy-mcp over stdio and checks its tool list
  cursor-deeplink.py              generates / checks the "Add to Cursor" link in docs
  export-hub.py                   renders docs/ as Galaxy Hub pages (content/tools/ai-agents/)
.github/
  ci-tools/                       CI-only npm deps (Pi loader check); not part of the Pi package
  workflows/validate.yml          per-harness CI jobs (see Testing)
  workflows/release.yml           packs the .mcpb and attaches it to tagged releases
  workflows/sync-skills.yml       weekly sync that opens a pull request

Why vendor the skills?

Each harness installs plugins by cloning or copying a directory. None of them is documented to follow git submodules (Pi runs a plain git clone), Claude Code discovers plugin skills only one level deep, Pi stops recursing at a skill root, and Pi can only install packages that carry a manifest. Vendoring a plain copy is the one layout that works everywhere; scripts/gen-manifests.py lists nested skills explicitly where a harness needs it. Each upstream tree is vendored into its own plugin rather than copying the repo root, which keeps upstream's using-vs-building split intact instead of registering both trees as one bundle. The Antigravity plugin.json at each plugin root must never carry an agent-plugins.org $schema: Codex and Cursor would then read it as an Agent Plugins manifest instead of their own (validate.py enforces this). The pinned upstream commit is recorded in each mirror's UPSTREAM.json, and the mirror plugins are versioned by sync date (YYYY.M.D).

Testing

Everything a harness does with these files is deterministic except the model itself, so CI checks the loaders rather than the models. validate.yml runs one job per harness on every push and pull request; none of them needs an API key or a login:

Job	What it proves
structure	validate.py: manifests parse, marketplaces match plugins/, every SKILL.md has frontmatter, nested skill lists and versions are regenerated, the Cursor deeplink in the docs is current
claude-code	claude plugin validate --strict on the marketplace and each plugin
codex	Codex's own plugin validator; then, with an isolated CODEX_HOME, installs all three plugins from the checkout as a local marketplace, checks codex mcp list shows galaxy, and reads the model-visible prompt with codex debug prompt-input to assert the skills are listed
antigravity	installs agy and runs agy plugin validate on each plugin
pi	loads every pi.skills root with Pi's own loadSkillsFromDir and checks the expected skills, no duplicates
mcp-smoke	starts uvx galaxy-mcp and the Claude Desktop bundle's uv run entry point over stdio, completes the MCP handshake and checks the core tools exist; with a GALAXY_API_KEY repository secret it also calls get_user against a live server
claude-desktop-bundle	mcpb validate + mcpb pack, uploaded as a build artifact

What CI does not cover, by design: whether a model picks the right skill. That needs a model call and a key. To add it later, the same commands used for manual checks work headlessly: claude --plugin-dir plugins/galaxy-skills -p "list your skills" with ANTHROPIC_API_KEY, and codex exec -m <model> with an OpenAI key. claude plugin eval can grade such runs if a suite is written under evals/.

Local equivalent of the whole matrix:

python3 scripts/validate.py
claude plugin validate --strict . && for p in plugins/*/; do claude plugin validate --strict "$p"; done
for p in plugins/*/; do agy plugin validate "$p"; done
python3 scripts/mcp_smoke.py
npx @anthropic-ai/mcpb validate bundles/claude-desktop/manifest.json

Maintaining

scripts/sync-skills.sh                        # main of both upstreams
FOUNDRY_REF=v0.4.0 scripts/sync-skills.sh     # pin a tag, branch or commit
python3 scripts/validate.py                   # what CI runs

Only plugins/galaxy-mcp/, docs/ and the hand-written manifest fields are edited here; changes to skill content belong upstream. The Galaxy Hub pages under /tools/ai-agents/ are generated from docs/ with scripts/export-hub.py <galaxy-hub>/content/tools/ai-agents; rerun it and open a hub pull request after changing a guide. A GitHub Action runs the sync every Monday and opens a pull request when anything changed.

Related

• galaxy-mcp, the MCP server
• galaxy-skills, skill sources
• foundry, the Workflow Foundry
• loom, a Pi-based Galaxy research harness

License

MIT (see LICENSE). Vendored skills keep their upstream licenses, copied alongside them.