Anthropic has open-sourced its full library of Claude Cowork plugins — 11 ready-made packages that transform Claude into specialized assistants for sales, marketing, finance, legal, customer support, product management, data, bio-research, and enterprise search. With 19,946 GitHub stars and an Apache 2.0 license, the repository is one of the clearest public signals of how Anthropic believes knowledge workers will actually deploy AI inside organizations in 2026: not as a single chatbot, but as a fleet of role-specific assistants that know which tools to use, which workflows to follow, and which slash commands to expose. ## What Claude Cowork Is Claude Cowork is Anthropic's platform for turning Claude into specialized assistants for different professional roles. Where consumer Claude is a general-purpose assistant, Cowork lets teams bundle three things into reusable packages: how they prefer work completed (skills), which tools and data Claude should reach into (connectors), and which slash commands to expose for explicit user-initiated workflows (commands). The plugins in this repository are Anthropic's reference implementations of each package — written in plain markdown and JSON, with no code required to install, customize, or extend. ## The 11 Plugins The set covers the major business functions in a mid-sized company. **Productivity** manages tasks and calendars across Slack and Notion. **Sales** handles prospect research and pipeline reviews via HubSpot and Clay. **Customer support** triages tickets and escalates issues through Intercom and HubSpot. **Product management** writes specifications and tracks competitive landscape via Figma and Amplitude. **Marketing** drafts content and plans campaigns using Canva and Ahrefs connectors. **Finance** handles reconciliation and statement preparation through data-warehouse connections. **Data** runs SQL queries and builds dashboards. **Legal** reviews contracts and tracks compliance across Box and Egnyte. **Enterprise search** unifies queries across email, chat, and documentation. **Bio-research** accelerates life-sciences work via PubMed and genomics databases. And **cowork-plugin-management** is the meta-plugin that lets organizations customize all of the above. ## Architecture: Markdown and JSON, Not Code The design decision that makes this repository interesting — and arguably the most consequential statement of Anthropic's product position — is that plugins are not code. Each plugin is a directory of markdown skill files, JSON command definitions, and MCP connector configurations. There is no SDK, no compilation step, no deployment pipeline. A skill is a markdown file that describes a domain expertise area and the workflow Claude should follow when that expertise is relevant; the model decides when to activate it. A command is a slash command users invoke explicitly, like `/finance:reconciliation`. A connector is an MCP server reference that links Claude to an external platform. This is a strong claim about who the plugin authors are supposed to be. The implicit answer is: domain experts, not engineers. A finance team lead writing a reconciliation workflow in markdown is the user the architecture is optimized for, and the file-based structure means a single pull request can extend or customize a plugin without coordination with an IT team. ## Customization Is the Real Value The out-of-the-box plugins are useful, but the documentation makes clear that the real value is in customization. Organizations are expected to modify connector configurations to point at their specific tools (their HubSpot, their data warehouse), embed company terminology and processes into the markdown skill files, and adjust workflows to match how their team actually operates rather than the textbook procedures encoded in the templates. The plugins are templates, in other words — designed to be forked and edited, not run as-is. ## Installation and Distribution For Claude.com users, installation happens directly through Cowork's plugin marketplace. For Claude Code, the workflow requires adding the `knowledge-work-plugins` marketplace before installing specific plugins. Once installed, relevant skills fire automatically when the model detects they're relevant; slash commands remain available throughout the session. The contribution model is a standard GitHub pull request against markdown files, which keeps the contribution barrier low for non-engineering organizations. ## Why This Repository Matters The interesting signal in the 19,946 stars is that the audience is not just Claude users — it's also a reference for everyone else building agent platforms. By open-sourcing the full plugin set, Anthropic has effectively published a recommended taxonomy for how role-specialized AI assistants should be structured: skills + commands + connectors, expressed in markdown and JSON, with MCP as the integration layer. Whether or not a given organization runs Claude Cowork, the architectural pattern is now in the open, and it's likely to shape how the rest of the agent-platform space organizes its own plugin systems in 2026 and beyond.