94 Claude Plugins Are Free, but the Word “Official” Deserves a Closer Look

Mika posted on X: Boris Cherny’s team at Anthropic shared 94 Claude plugins and skills, all free. The number is striking. Fifty-five are for Claude Code, 20 target real roles such as sales, legal, finance, and marketing, and 19 are skills for things like Excel, PowerPoint, PDF, and Word. Mika also added a comparison: these repositories total about 269,000 GitHub stars, and quotes for having someone build one of them range from $300 to $1,500.

Let’s unpack the number first. The 94 is the total of plugins and skills, not 94 products written by Anthropic itself. The official plugin directory separates internal plugins from external ones, and external plugins come from partners and the community. Commenter Kuz noted that most of the official marketplace comes from partners and tool vendors; Anthropic’s own collection is smaller, such as code-review, commit-commands, and feature-dev, and these have already been added automatically in Claude Code for several months. Mika says 20 target real roles, while the knowledge-work plugins repository currently lists 11, covering sales, legal, finance, marketing, product, customer support, data, enterprise search, and more. The definitions differ, so check the source before installing.

What are plugins? They are file-based configurations, usually containing plugin.json, .mcp.json, commands, agents, and skills. They package tool connections, slash commands, and workflow instructions. They do not change the model itself; what they change is which tools Claude can access and what workflow it follows. So the real risk is permissions. Commenter Kagi said that before installing, open the files and see what it can run—plugins can execute commands. Mika replied that permissions deserve as much attention as the demo. One commenter joked that a plugin for chasing overdue payments might be more popular than one for replying to emails; someone else asked, once the draft is written, who presses send? That question matters more than the number of plugins.

How to install. In Claude Code, type /plugin, browse Discover, or directly run /plugin install {plugin-name}@claude-plugins-official. For knowledge-work plugins, first add the marketplace: claude plugin marketplace add anthropics/knowledge-work-plugins, then install an individual one, for example claude plugin install sales@knowledge-work-plugins. For the Skills repository, first run /plugin marketplace add anthropics/skills, then install document-skills@anthropic-agent-skills or example-skills@anthropic-agent-skills.

A more stable starting point. Mika’s other Claude Duet tutorial gives a minimal structure: one folder; CLAUDE.md states identity, way of working, and prohibitions; memory/people and memory/projects store people and projects; .claude/commands holds /morning-brief; and .claude/settings.json spells out allow, ask, and deny. For unattended operation, use claude -p to run scheduled tasks, but the machine has to stay awake. This structure explains the threshold for proactive agents better than installing 94 plugins: permission boundaries live in configuration files, not in demo videos.

Ninety-four plugins are too many, and the screening cost is real. The plugins are free; what is not free is your time spent reading their permissions, configuring their context, and testing whether their output is reliable. A safer approach is to pick one task that repeats every week—meeting prep, weekly reports, initial contract review, inbox triage, anything like that. Start read-only; don’t let the plugin send emails, make payments, or delete files yet. Run it for a week and see whether the output can be traced back to its sources. If not, delete it. Once it saves you that repeated hour each week, consider a second one.

Links:
Official plugin directory
Knowledge-work plugins
Skills repository
Mika’s Claude Duet tutorial
Original tweet

If you have already installed one, look at its plugin.json and .mcp.json first, then decide whether to keep it.