wshobson/agents

A single-source marketplace of agents, skills, and commands for Claude Code that also generates native artifacts for Codex CLI, Cursor, OpenCode, Gemini CLI, and GitHub Copilot.

What it is

wshobson/agents is an agentic plugin marketplace that packages reusable workflow building blocks — agents, skills, commands, and orchestrators — as Markdown from one source of truth in plugins/. From that source it emits harness-native artifacts for five additional harnesses rather than lowest-common-denominator translations, so the same content works idiomatically across Claude Code, OpenAI Codex CLI, Cursor, OpenCode, Gemini CLI, and GitHub Copilot. It is aimed at developers who want production-ready domain experts and slash commands they can install per-plugin instead of adopting a whole monolith.

Key features

A catalog of 94 plugins, 203 agents, 175 skills, 109 commands, and 16 multi-agent orchestrators, per the README's own counts.
Granular, single-purpose plugins; installing a plugin loads only its components into context, not the whole marketplace.
Multi-harness generation from one Markdown source, with per-harness adapters (for example .codex/, .cursor/rules/, .opencode/, Gemini TOML, .copilot/) driven by make generate-all.
A tiered model strategy that maps each agent to a model tier (Fable 5, Opus, inherit, Sonnet, Haiku) by task type.
plugin-eval, a three-layer quality framework: deterministic static analysis, an LLM-judge across four dimensions, and Monte Carlo reliability runs.
Make-based tooling for structural validation and drift/dead-link/cap detection (make validate, make garden).

Tech stack

Primary language reported as Python, used by the plugin-eval framework and run via uv (uv run plugin-eval ...).
Content authored as Markdown (agents, skills as SKILL.md, commands) as the source of truth.
JSON marketplace/plugin registries per harness (for example .claude-plugin/marketplace.json, .agents/plugins/marketplace.json, .cursor-plugin/), and TOML output for Gemini CLI.
A Makefile orchestrates generation, validation, and installation targets.
MIT licensed.

When to reach for it

You use Claude Code and want a large library of ready-made domain agents, skills, and slash commands.
You work across more than one agentic harness and want the same content available natively in each.
You want to install narrowly scoped plugins so only relevant components enter the model's context.
You want a repeatable way to measure and certify the quality of agents and skills before shipping them.

When not to reach for it

You need runtime multi-agent orchestration logic; this is primarily a content marketplace, not an execution engine.
You want a plug-and-play binary; Gemini and OpenCode install via clone plus make generate, and some trees are gitignored and must be built.
You only ever use one harness and prefer that harness's own first-party plugin ecosystem.
You need harness features beyond the documented capability matrix; some deltas (for example an 8 KB skill cap on Codex) are respected by trimming or remapping content.

Maturity signal

The project has over 38,000 stars gathered since its creation in July 2025, a most-recent push in July 2026, and only 7 open issues, which together read as very actively maintained with a well-tended issue tracker. The MIT license and single-source, multi-harness architecture suggest a project built for ongoing extension rather than a one-off dump, and the presence of a dedicated evaluation framework and CI workflows (validate, code-quality, eval-report) reinforces active quality control.

Alternatives

Yeachan-Heo/oh-my-claudecode — use it instead when you want a runtime multi-agent orchestration system rather than a catalog of installable content.
Claude Code's native subagents and plugins — use them instead when you want to stay entirely within first-party tooling and define agents yourself.
Community "awesome" Claude Code lists — use them instead when you want a curated index of resources rather than installable, harness-native plugins.

Notes

The five non-Claude harnesses receive purpose-built artifacts, not shared translations: Codex and Cursor install from committed registries that point back at the source plugins/, while Gemini and OpenCode install through clone-plus-generate with their transformed trees gitignored. The repo also includes an external memory integration, Pensyve, wired in as a git-subdir entry with separate upstream integrations for each supported harness.

Wiki: wshobson/agents