Agentic Plugin Marketplace (wshobson/agents)

Why it's ranked here

The case rests on breadth plus plumbing. MIT licensed, installable in Claude Code with two commands, and generated for six other agent tools from the same source, including Codex CLI, Cursor, OpenCode and Copilot. The repository also ships its own checks: a structural validator, a drift detector for dead links and oversize skills, and a pytest suite, all wired into CI per its architecture notes. Plenty of prompt collections exist; fewer come with a build step that keeps several harnesses in sync and refuses to delete files outside the repository or a temp directory.

What's good

The bundle design keeps context small: you install the two or three bundles for your domain and nothing else is loaded. Skills follow a three-tier layout where only the name and trigger description are always loaded and the body arrives when activated. The generator is careful about destruction: it refuses a clean of one bundle because that would wipe every other bundle's output, and it only prunes inside the folders it owns, leaving a developer's own Pi settings alone. The quality-scoring tool is unusually candid, labelling its two model-based layers experimental and listing their known limits in its own documentation.

Tradeoffs

The headline counts do not agree across the repository's own files: the README says 94 bundles and 184 skills, the marketplace manifest says 92 and 181, and the architecture guide says 183 skills. Two architecture documents also describe the model tiers differently, one with four tiers and one with five. Only Claude Code, Codex and Cursor install from the committed registries; Antigravity, OpenCode and Pi need a clone plus a local generate step. The quality scores are mostly a lint: bundle-level badges come from the static layer alone, and the model-judged layers are unvalidated against human labels, by the project's own account.

How to use it well

Treat it as a menu, not a download. Browse the catalogue, install the few bundles matching your stack, and let the agent pick skills up by their descriptions. If you only want the knowledge packs, the skills-only installers pull individual skills into any supported agent without the personas or commands. Teams running several agent tools get the most from it, since one Markdown source feeds all of them. It is not a runtime or framework: it supplies prompts and instructions, so output quality still depends on the model you point it at. Read the quality badges as structure checks, not proof that a skill helps.

Observed

License

MIT

Content format

Markdown agents, commands and skills with YAML frontmatter

Generator language

Python (tools/generate.py)

Primary install surface

Claude Code plugin marketplace (/plugin marketplace add wshobson/agents)

Other harnesses generated

Codex CLI, Cursor, OpenCode, Antigravity CLI, GitHub Copilot, Pi

Skills-only install

gh skill install and npx skills add, no clone required

Build and checks

make generate, make validate, make garden, make test (pytest)

Quality tool requirement

Python 3.12 or newer, managed with uv

Count consistency

README and marketplace manifest state different plugin and skill totals