wshobson/agents: 94 Plugins, One Marketplace, Seven Agent Harnesses

You find a great plugin collection, install it, then switch from Claude Code to Codex for a week — and the collection stays behind. wshobson/agents, a 40k-star marketplace, is built against that problem: one source of truth under plugins/, from which seven supported harnesses each receive artifacts in their own idiom.

Why This Skill Matters

The catalog is large: 94 plugins (92 local plus 2 external via git-subdir), 202 domain-expert agents, 183 skills with progressive disclosure, 105 slash commands, and 16 multi-agent orchestrators. The README's positioning note matters as much as the totals: each adapter emits harness-native artifacts, not lowest-common-denominator translations. Codex gets commands converted into skills under its 8 KB skill cap, Cursor gets a thin marketplace plus curated rules that reuse .claude/, and OpenCode derives its permission block from the source tools: allowlist.

Plugins stay isolated: installing one loads only its components into context, not the whole marketplace. The python-development plugin, for example, carries 3 Python agents (python-pro, django-pro, fastapi-pro), 1 scaffolding command, and 16 specialized skills covering async, testing, and packaging.

Installation

Inside Claude Code:

/plugin marketplace add wshobson/agents
/plugin install python-development

Codex CLI installs from a committed registry:

npx codex-marketplace add wshobson/agents

Only want the skills, in any agent? The skills-only route reads plugins/*/skills/ straight from GitHub — no clone, no marketplace, no generate step:

gh skill install wshobson/agents
npx skills add wshobson/agents --skill python-testing-patterns

OpenCode, the Antigravity CLI, and Pi install via clone plus generate:

gh repo clone wshobson/agents ~/agents && cd ~/agents
make install-opencode

Real Workflow: Scope an Install to What You Need

1. Pick the narrowest plugin that matches the work — a Python backend project wants python-development, not five adjacent plugins.
2. Install it in Claude Code with the two /plugin commands above.
3. Open a task in your repo. The plugin's agents, command, and skills auto-discover from the directory structure, and skills load when activated — the rest of the marketplace never enters your context.

Each agent also carries a model tier from the README's strategy table: Opus for architecture, security, and code review; Sonnet for docs, testing, and debugging; Haiku for fast operational tasks like deployment and content. Tier 0 (Fable 5) is opt-in for the longest autonomous runs.

Real Workflow: Certify a Skill Before You Trust It

The repo ships a three-layer evaluation framework, plugin-eval, for measuring plugin and skill quality:

uv run plugin-eval score path/to/skill --depth quick
uv run plugin-eval certify path/to/skill

Static analysis is deterministic and free, the LLM judge scores four dimensions in about 30 seconds, and the Monte Carlo layer runs 50–100 simulated executions to measure reliability. Run it on skills you author yourself before publishing them anywhere.

Tips

• make validate runs structural checks across generated outputs; make garden detects drift and dead links.
• Browse before installing: the full catalogs live in docs/plugins.md, docs/agents.md, and docs/agent-skills.md.
• Two external entries ship as git-subdir: the Pensyve memory integration and a HOL Guard security entry for Claude Code, whose reviewed payload is pinned to a commit and requires user approval.
• On Codex, adding the marketplace and installing a plugin are two separate steps — pair them, as the README's quick start does.

When Not to Use This

If your work lives in one narrow stack, a 94-plugin marketplace is more surface than you need — take the skills-only route and pull the two or three skills you will actually use. And if you are on OpenCode, Antigravity, or Pi, expect the clone-plus-make path rather than a one-line marketplace add.

See the leaderboard for more skills.