Claude Plugins Official is Anthropic's curated, in-repository marketplace of high-quality plugins for Claude Code, the official CLI agent for Anthropic's Claude. The repository, which has accumulated 20,800+ stars and 2,549 forks since opening, serves as the canonical source agents and developers reach for when they run `/plugin install` inside Claude Code. It hosts both `/plugins` developed and maintained by Anthropic itself and an `/external_plugins` directory for third-party submissions that have passed quality and security review. The directory is reachable directly from Claude Code via `/plugin > Discover`, giving the agent a built-in path to extend itself with new commands, agents, skills, and MCP servers. ## A Marketplace That Lives in a Git Repo Unlike traditional plugin stores backed by proprietary registries, Claude Plugins Official is just a public GitHub repository. Each plugin is a directory containing a `.claude-plugin/plugin.json` manifest, optional MCP server configuration, slash commands, agent definitions, skills, and a README. Claude Code knows how to read this layout, so installation is `git clone` plus path resolution rather than a remote download protocol. The transparency this creates is a feature: every byte of every plugin is auditable, diff-reviewable, and historically pinned by commit, which matters because plugins can register MCP servers and slash commands that act with the agent's permissions. ## Two Tiers: Internal and External The `/plugins` directory holds plugins built and operated by Anthropic team members. These typically wrap official Anthropic infrastructure, ship updates in lockstep with Claude Code releases, and carry the highest trust expectation. The `/external_plugins` directory hosts partner and community submissions that have cleared Anthropic's quality and security gates before merge. The distinction is explicit in the directory layout, so users can choose how much external code they are comfortable executing inside their agent. ## Standard Plugin Anatomy Every plugin in the directory follows the same minimal layout: a `.claude-plugin/plugin.json` manifest declaring name, version, and entry points; an optional `.mcp.json` for tool-server registration; a `commands/` directory of slash commands; an `agents/` directory of sub-agent definitions; a `skills/` directory of executable skills; and a `README.md` documenting usage. This uniform structure means plugin authors can copy any existing entry as a template, and Claude Code can load any of them without per-plugin shimming. The schema is the same one used by `/plugin install` outside this directory, so plugins built here transfer cleanly to private marketplaces. ## Installation Without Leaving the Agent Installing a plugin requires neither a browser nor a separate package manager. Inside Claude Code, running `/plugin install <name>@claude-plugins-official` resolves the entry from this repository and loads it into the running session. The GUI alternative `/plugin > Discover` lists every plugin in the directory with its description, letting users browse before installing. New slash commands, agents, or skills become available immediately, and uninstalling reverses the operation cleanly. ## Why It Matters A curated, official directory anchors the rapidly fragmenting plugin ecosystem. Plugin marketplaces have historically suffered from either over-permissive registries that ship malware (npm, PyPI typosquats) or walled gardens that throttle developer velocity. By using a public Git repo with explicit quality gates and Anthropic-operated curation, Claude Plugins Official tries to thread the needle: anyone can submit via the [plugin directory submission form](https://clau.de/plugin-directory-submission), but every entry is human-reviewed before merge, and the repository's commit history is the audit trail. ## Submission and Curation External submissions go through a documented review process that checks both quality (does the plugin actually deliver the value its README claims) and security (does it execute unexpected code, exfiltrate data, or register opaque MCP servers). Internal plugins, by contrast, are developed inside Anthropic and updated through the standard pull-request workflow on this repository. Anthropic explicitly warns that even with curation, users should verify they trust a plugin before installation because the framework cannot fully audit every MCP server or external file referenced. ## Limitations The directory ships no license file at the repository root, which means individual plugins carry their own licensing terms and users must check each before reuse beyond personal installation. Curation is human-paced, so the directory is intentionally not a high-volume firehose like npm. And because the repository is the source of truth, plugins published elsewhere are not discoverable via `/plugin > Discover` unless mirrored here.