Agentic Plugin Marketplace: Claude Code plugin marketplace
/agents
Agentic Plugin Marketplace solves a real problem: installing one plugin loads only that plugin's agents and skills, not the whole 202-agent catalog. That per-plugin isolation beats a single giant prompt library on context cost alone. The weak spot is harness parity — Gemini CLI and OpenCode need a manual clone-and-generate step, and Copilot gets no single install command at all.
What is the Agentic Plugin Marketplace?
Agentic Plugin Marketplace is a catalog of 91 installable plugins built from 202 agents, 180 skills, and 105 commands, written once in Markdown and generated natively for Claude Code, Codex CLI, Cursor, OpenCode, Gemini CLI, and Copilot. Each plugin bundles only the pieces it needs — installing python-development pulls in three agents and sixteen skills, not the entire marketplace.
What the marketplace includes
- ✓91 installable plugins (90 local plus 1 external via git-subdir) covering areas like Python development, security, and infrastructure
- ✓202 domain-expert agents spanning architecture, languages, infra, security, data, ML, docs, business, and SEO
- ✓180 skills that load only when a plugin activates them, using progressive disclosure instead of dumping everything into context
- ✓105 slash commands for scaffolding, security scans, test generation, and infrastructure setup
- ✓16 orchestrators that coordinate multiple agents for workflows like full-stack builds, security reviews, ML pipelines, and incident response
- ✓A five-tier model strategy: Fable 5 for long-horizon work, Opus for production-critical review, Sonnet for docs and testing, Haiku for fast operational tasks
- ✓The plugin-eval framework, which scores plugins with a static check, an LLM judge, and Monte Carlo simulation runs
Installing for your AI coding tool
In Claude Code, run `/plugin marketplace add wshobson/agents`, then `/plugin install python-development` — or any of the 91 plugins. Codex CLI installs from the same source with `npx codex-marketplace add wshobson/agents`, then plugins go in one at a time. Cursor reads the committed `.cursor-plugin/` registry — add the marketplace, then run `/plugin install <name>`. Gemini CLI and OpenCode skip the registry entirely: clone the repo, then run `make generate HARNESS=gemini && gemini extensions install .` for Gemini, or `make install-opencode` for OpenCode. Copilot gets no single command. Its output format (`.copilot/agents/`, `.copilot/skills/`, `.copilot/commands/`) is documented, but the install path itself is not clearly documented in the README.
Orchestrators and multi-agent workflows
- •Install python-development to pull in python-pro, django-pro, and fastapi-pro agents plus the async, testing, and packaging skills for one Python project
- •Run one of the 16 orchestrators to coordinate a full-stack build or security review across several agents instead of prompting each one separately
- •Route architecture and security review to the Opus tier while SEO and deployment tasks run cheaper on Haiku
- •Score a custom skill with plugin-eval before shipping it — static check first, then the LLM judge, then a Monte Carlo run if it needs a reliability number
Strengths
- ✓Installing a plugin loads only that plugin's agents, commands, and skills — not the full 202-agent catalog
- ✓One Markdown source generates harness-native artifacts per tool instead of a lowest-common-denominator translation
- ✓The tiered model strategy matches Fable 5, Opus, Sonnet, or Haiku to how critical or long-running the task is
- ✓plugin-eval gives you a static check, an LLM judge, and Monte Carlo runs to actually score a plugin instead of trusting it blind
- ✓16 orchestrators handle multi-agent coordination instead of leaving you to sequence agents by hand
Limitations and considerations
- △Copilot's install path isn't spelled out with a one-line command in the README the way the other five harnesses are
- △Gemini CLI and OpenCode need a manual clone-and-generate step instead of installing straight from a committed registry
- △The generated Gemini and OpenCode trees are gitignored, so you re-run `make generate` locally rather than pulling a maintained artifact
- △Tier 0 (Fable 5) is described as opt-in and premium cost, without further detail on what that costs
- △With 202 agents and 180 skills in the catalog, finding the right plugin for a narrow task means reading docs/plugins.md rather than browsing in-tool
Alternatives for AI coding assistant extensibility
Frequently asked questions
In Claude Code, run `/plugin marketplace add wshobson/agents` to register the marketplace, then `/plugin install <plugin-name>` — for example `/plugin install python-development` — to add any of the 91 available plugins.
Agentic Plugin Marketplace targets Claude Code, Codex CLI, Cursor, OpenCode, Gemini CLI, and GitHub Copilot, generating harness-native agents, commands, and skills for each one from a single Markdown source.
Agentic Plugin Marketplace generates the same plugin content into harness-specific formats for Claude Code, Cursor, and Codex CLI, so the same plugins work across all three — though each install path differs: Claude Code and Codex CLI use committed registries, Cursor reads the `.cursor-plugin/` registry alongside the source.
Agentic Plugin Marketplace assigns each agent a model tier: Tier 0 (Fable 5) for long-horizon autonomous work, Tier 1 (Opus) for architecture, security, and production-critical review, Tier 2 for user-chosen models, Tier 3 (Sonnet) for docs and testing, and Tier 4 (Haiku) for fast operational tasks.
plugin-eval is Agentic Plugin Marketplace's three-layer evaluation tool: a static structural check, an LLM judge scoring four dimensions with Haiku and Sonnet, and a Monte Carlo run simulating 50 to 100 executions to measure reliability.
The problem it solves
Claude Code plugins, agents, and skills are usually written once for Claude Code and never touch Codex CLI, Cursor, Gemini CLI, or Copilot — every team building agentic tooling either commits to one harness or maintains separate copies by hand. Agentic Plugin Marketplace keeps a single Markdown source in `plugins/` and generates harness-native artifacts for each tool instead of asking you to port agent definitions yourself.
Who should try it — and who should skip
Teams already standardized on Claude Code get the cleanest experience: install one plugin, get exactly its agents and skills, and nothing extra loaded into context. Developers who split time across Gemini CLI or OpenCode take on more setup — both need a clone-and-generate step before anything shows up in the host. If you only need one or two agents for a narrow task, docs/plugins.md and a single plugin install beat pulling in the whole catalog.
Related repositories
Still deciding about agents?
One click hands the question to an AI along with this page — see what it says about agents.