Skip to main content
Everything you can add to your AI agent, in one place - ranked by what people actually install. Scan the list, or head into a category when you know what you’re looking for.

​
Plugins, skills, and MCP servers - what’s the difference?

Every entry in this directory is something you add to an AI agent - think of each one as an extension that gives the agent new tools, knowledge, or a working process. The same extension often ships through more than one route: as a plugin for Claude or Codex, as a generic MCP server, as a standalone skill. So we list each one once and show all of its routes, rather than listing the same thing four times under four names.
  • Plugin - a packaged bundle installed inside the agent: commands, skills, and MCP connections in one install, distributed through marketplaces.
  • Skill - a folder of instructions the agent reads and applies when relevant. The lightest route, and the most portable: the same SKILL.md format now works across many agents, with .agents/skills/ emerging as the shared vendor-neutral location.
  • MCP server - a running service the agent connects to for live tools and data, using the open Model Context Protocol. Works in any MCP-capable agent.
  • CLI - a command-line tool the agent runs in the terminal. Works with any agent that can run shell commands.

​
What each agent calls them - and how to install

For any agent not listed here: if it speaks MCP, add the extension’s MCP server to its config; if it loads skills or instruction files, copy the skill folder into wherever it reads them from. Those two routes are the portable ones - they work the same everywhere.

​
How we estimate installs

No single authoritative source exists for how widely an agent extension is deployed. Vendors rarely publish telemetry and most tools ship through several channels at once, so we analyze thousands of public signals across multiple sources and reconcile them into a single estimate. Direct installation counters receive the greatest weight. These can include Claude plugin installs from the exact listing. skills.sh installs from the exact skill or collection page, with skill-level and collection-level counts kept separate. Visual Studio Marketplace cumulative installs, counted only where the extension materially provides the tool rather than merely mentioning it. Where these exist, they anchor the estimate and always take priority. Distribution data helps when no authoritative counter exists. We examine downloads for exact, verified npm and PyPI packages and pulls for Docker images that serve as a tool’s primary distribution route. Download activity is not equivalent to installed users: it can include updates, automated builds, mirrors, and repeated downloads. We therefore normalize it by channel and calibrate it against extensions for which both distribution data and direct installation counters are available. Public implementation evidence. We search public code for exact package names, server identifiers, and configuration entries such as those found in .mcp.json files. These matches can indicate that developers have implemented or evaluated a tool, but they do not prove that every configuration remains active, and they exclude usage in private repositories. Attention metrics are used as supporting evidence. GitHub stars, repository activity, search demand, documentation traffic, and public discussion can help validate the general order of magnitude or reveal adoption that other counters miss. They are not converted directly into installations. Repository signals are considered only when the repository is clearly associated with the tool’s primary development or distribution. Every source is verified and overlaps are reconciled. We tie evidence to the exact listing, package, image, or repository and exclude similarly named forks, mirrors, examples, and unrelated integrations. We also account for wrappers around existing packages, monorepos that publish the same tool through multiple channels, and traffic likely caused by continuous integration or automated systems. These estimates are intended to show relative market adoption, which we use as one directional input in our rankings: an extension with broader adoption may be better maintained, and more likely to be relevant to a larger number of users, but popularity alone does not make it the best choice. The right extension for you depends on your actual need and the tools you already run; each listing lays out capabilities, requirements, and integrations so you can make that call.