wshobson/agents

What It Is

wshobson/agents is the largest known plugin ecosystem for Claude Code — and, as of 2026, no longer only for Claude Code. The project rebranded as a "multi-harness agentic plugin marketplace": 192 specialized AI agents across 84 focused plugins, covering architecture, languages, infrastructure, security, data/AI, documentation, business operations, and SEO, all authored once in Markdown and shipped as harness-native artifacts to Codex CLI, Cursor, OpenCode, Gemini CLI, and GitHub Copilot.^[1]

The system follows a granular, composable design — each plugin is isolated with its own agents, commands, and skills, and installing a plugin loads only its components into context, not the whole marketplace.^[1]

It is free, MIT-licensed open source; the only cost is the model usage of whichever harness you run it in.^[2]

Architecture

As of June 2026:^[1]

• 84 Plugins — Single-purpose, minimal token footprint (82 local + 2 external via git-subdir). Up from 72 in February 2026.
• 192 Agents — Domain experts spanning full-stack development, DevOps, security, data, and ML. Up from 112.
• 156 Skills — Modular knowledge packages with progressive disclosure for token efficiency. Up from 146.
• 102 Commands — Slash commands for scaffolding, security scans, test generation, and infrastructure setup (formerly billed as 79 "tools").
• 16 Orchestrators — Multi-agent coordination for complex workflows (full-stack dev, security hardening, ML pipelines, incident response). Unchanged.

Key Features

Multi-Harness Support (New in 2026)

The biggest change since this profile's first edition: one source-of-truth (plugins/) now generates idiomatic artifacts for six harnesses. Claude Code remains native; Codex CLI gets committed registries with its 8 KB skill cap respected; Cursor gets a thin marketplace plus curated rules; OpenCode gets permission blocks derived from tool allowlists; Gemini CLI gets TOML-native skills and subagents; Copilot gets Markdown agent profiles with models mapped to the GPT-5 family.^[3]

Agent Teams (Parallel Workflows)

The Agent Teams plugin enables Claude Code's multi-agent parallelism — team presets for review, debugging, feature work, and security, including parallel code review across security, performance, and architecture dimensions and hypothesis-driven debugging with multiple parallel investigations.

plugin-eval (Quality Certification, New in 2026)

A three-layer evaluation framework for measuring and certifying plugin/skill quality: deterministic static analysis (under 2 seconds, free), an LLM judge scoring four dimensions, and Monte Carlo statistical reliability testing via 50–100 simulated runs.^[1]

Four-Tier Model Strategy

Updated for Opus 4.7: tier 1 (Opus 4.7) handles architecture, security, and production-critical work; tier 2 inherits the user's model choice; tier 3 (Sonnet) covers docs, testing, and debugging; tier 4 (Haiku) runs fast operational tasks.^[1] The February 2026 edition of this profile described a three-tier Opus 4.6/Sonnet 4.6/Haiku 4.5 scheme.

Why It Matters

This project demonstrates the emerging pattern of plugin marketplaces for coding agents — and is now a live experiment in whether one skill/agent corpus can serve every major harness. Rather than monolithic agent systems, wshobson/agents bets on composability: install only what you need, keep token budgets tight, and orchestrate multiple specialized agents for complex tasks.

With 36.6K GitHub stars and ~3.9K forks as of June 2026 (up from 29K stars in February), it remains one of the most popular open-source projects in the Claude Code ecosystem and a reference implementation for plugin-based agent architectures.^[2] Its skills are also distributed through the Smithery registry.^[4]

Cautions

• Single-maintainer concentration. Despite the star count, this is essentially one person's (Will Hobson's) curation. There are no versioned GitHub releases — you track main, and what main contains can change substantially (agent count went 112 → 192 in four months).^[2]
• Decision paralysis. 84 plugins and 192 agents is a catalog, not a starting point. Our agentic skills frameworks comparison flagged plugin selection and configuration as an hours-long onboarding cost.
• Suspiciously quiet issue tracker. Around a dozen open issues for a 36K-star repo suggests under-reporting rather than flawless software — most users likely fork or fix locally rather than file.^[2]
• Uneven multi-harness ergonomics. Codex and Cursor install from committed registries, but Gemini CLI and OpenCode require clone + make generate; generated trees are gitignored.^[3]
• Quality varies by prompt corpus. These are prompt-engineering artifacts, not tested code; the new plugin-eval framework exists precisely because quality across 156 skills is hard to guarantee.

What Developers Say

Substantive public reviews are scarcer than the star count would suggest — most commentary is in listicles rather than first-hand reports. The first-hand accounts that do exist are positive:

"Treat prompts as source code - edit them in files, use @ notation to bring them into the console. Use Claude to generate its own prompts - https://github.com/wshobson/commands/ and https://github.com/wshobson/agents/ are very handy, they include a prompt-engineer persona." — jaggederest, Hacker News^[5]

"The only Skills you need. Clone wshobson/agents, copy the tools folder to ~/.codex/prompts. Try 'explain this codebase' raw, then run the same prompt with /prompts:code-explain. You'll see a huge difference in the quality of your output." — yummyelephant8, Hacker News, March 2026, describing a Codex CLI workflow^[6]

The second quote is notable: developers were porting the collection to Codex by hand before the project shipped official multi-harness support.

Bottom Line

Recommended if you live in Claude Code (or Codex CLI/Cursor) and want a curated, composable library of domain-expert agents and skills without writing your own — it's free, MIT-licensed, actively maintained (last push June 8, 2026), and the granular plugin design genuinely limits context bloat.^[2]

Not recommended as a turnkey orchestration platform or for teams that need versioned, stable dependencies — there are no releases, the catalog churns fast, and orchestration plugins still demand significant configuration and debugging.

Outlook: the 2026 multi-harness pivot is the story to watch. If one Markdown corpus can stay idiomatic across Claude Code, Codex, Cursor, OpenCode, Gemini, and Copilot, wshobson/agents becomes infrastructure for the whole coding-agent ecosystem rather than a Claude Code accessory; if the adapters drift, it risks spreading a single maintainer across six moving targets.