Refactor repo as agent-skills

[bentsherman]

This PR refactors the repo to a generic "agent-skills" instead of a Claude plugin

• Rename project to agent-skills
• Use npx skills to install skills for any coding agent

See also:

• https://github.com/mattpocock/skills
• https://www.skills.sh/docs

[ewels]

Nice 👍🏻

Addressed

[pditommaso]

Thanks for putting this together — broader cross-agent reach is the right goal. A few drawbacks worth weighing before going in this direction:

No MCP bundling for launch-workflow. The current Claude plugin manifest declares the Seqera MCP server and prompts the user to approve it during install — that's how launch-workflow becomes usable out-of-the-box. The new README replaces "approve the MCP server when prompted" with "make sure the Seqera MCP server is configured for your agent" but doesn't say how. First-time launch-workflow users will hit a missing-tool error with no recovery path in the README.

No marketplace visibility in Claude Code. Today the plugin is discoverable from inside Claude via /plugin marketplace. With npx skills replacing the marketplace path, Claude users have to leave the agent to install, and we lose the marketplace listing entirely.

No auto-update. Claude Code marketplace plugins check on launch and pull updates automatically. npx skills@latest add … is a one-shot copy — the @latest refers to the installer version, not the skill. After install, the SKILL.md files become static; users only see fixes when they manually re-run the install command. For an actively-iterating project that's a real operational regression — e.g., the 26.04.0 → 26.04 change in this PR would land instantly for marketplace users but only when skills.sh users remember to re-install.

Why 26.04.0 → 26.04? nextflow -version reports the full triplet (26.04.0, 25.10.1). The new wording in install-nextflow ("older than 26.04") sets up an awkward comparison against the reported 26.04.0, and "26.04 — first release of April 2026" is technically inaccurate since the actual tag is 26.04.0. Curious what motivated the change?

Alternative. I opened #7 with a different shape that gets the cross-agent reach without losing the Claude UX:

• Keep the Claude Code plugin marketplace as the native install path — auto-update, MCP bundling, and marketplace listing all preserved.
• Ship an install.sh for Codex, Cursor, Gemini CLI, GitHub Copilot, Goose, Windsurf, and OpenCode. Per-agent destination paths verified against each agent's own docs.
• No third-party hosted infra (no skills.sh dependency).

Note: #7 intentionally keeps the repo name as claude-plugin for now. Once we consolidate on a direction, we can rename the repo (and update the README + plugin.json URLs) in a follow-up.

See #7 — worth comparing the two shapes before we pick a direction.

[bentsherman]

I'm not sure how to proceed here. AFAIK the npx skills route is the standard convention for installing skills in an agent-agnostic manner

I will address what I can but I must leave the rest to @ewels and @edmundmiller (especially the points about MCP and auto-updates)

Ship an install.sh for Codex, Cursor, ...

The install.sh in #7 seems like a downgrade compared to npx skills. Users will already have npx, they won't want to clone this repo just to run a script to install the skills

No marketplace visibility in Claude Code

The repo can still function as a CC plugin, which is why I didn't remove the .claude/plugin.json in this PR. We can document it as a parallel install path if you want. You shouldn't lose visibility in the CC marketplace

Why 26.04.0 → 26.04?

Not really a big deal either way. When referring to major versions, I prefer to omit the patch number since it's not relevant. The agent should just use the latest patch release of the major version.

[bentsherman]

Talking to Edmund, I think I will refactor the README to provide focused install paths for the most common harnesses:

https://github.com/edmundmiller/nf-core-skills

Instead of using npx skills

[ewels]

Yup, this echoes a discussion we had on Slack (somewhere). If we want to then I think we could provide all things. This is what the folks do at impeccable: https://github.com/pbakaus/impeccable

I'd recommend pointing Claude or similar at the repo to explain. It looks complicated as hell but it's mostly as there is loads of duplication. Automation takes a key set of files and meta and then copies it out into all the different formats. Then you can use that one repo for npx skills and native harnesses and basically all the things. Makes the repo footprint large but it's all automated so the complexity isn't as bad as it looks.

As is usually my style, I would advocate for this approach (maybe not all those agents though, could pick the top 3 or 4). Provide all the installation options at once, so people can pick their preferred method..

[pditommaso]

I had looked at impeccable and if I'm understanding correctly they use a kind of templating + build system that looks overkill to me at least at this stage.

Not getting a simple install script is a "downgrade". We are talking of literally copying 4 markdown files, don't see where the downgrade is.

Anyhow I'm ok with any solution provided we don't break the ability to use directly with claude (ie. marketplace/github installation, mcp support, etc)

[bentsherman]

Edmund also shared this: https://github.blog/changelog/2026-04-16-manage-agent-skills-with-github-cli/

So users could just install using the gh CLI:

gh skill install nextflow-io/skills

Seems like the same experience as npx skills, but I suspect our users are more likely to already have gh installed

[bentsherman]

I updated the PR based on the discussion. The README now provides two install paths:

1. Claude Code plugin: specific to CC, Seqera MCP included, auto-updates
2. GitHub CLI: harness-agnostic, Seqera MCP not included, update manually with gh skill update

This keeps the benefits of the CC plugin while providing a simple install path for other harnesses. Most of our users will already have gh installed

@pditommaso ready for review

[bentsherman]

By the way, it looks like Seqera MCP is only used by launch-workflow to select a compute environment. But you can do that with nextflow auth login as well. I'm not sure how the agent handles the interactive CLI flow. But if we can use that here then we might not need the MCP at all