In This Article
9 sectionsQuick answer
Claude Cowork plugins bundle skills, connectors, agents, and hooks into one package. Learn what they do, how to install them, and how to stay safe.
Claude Cowork plugins are packaged bundles that extend what Claude can do inside Cowork. One plugin can add skills, MCP connectors, subagents, slash commands, and hooks in a single install. They come from a marketplace, from your organization, or from a file you upload, and each component can be switched on or off.
Cowork plugin limits and the install flow verified 31 July 2026 against Anthropic's plugins documentation.
Cowork is Anthropic's agentic workspace. It runs the same agentic architecture that powers Claude Code, except it lives inside Claude Desktop rather than a terminal, so you describe an outcome and come back later to finished work. If the product itself is new to you, our guide to what Claude Cowork is covers the ground floor, and the breakdown of Cowork versus Claude Code explains why one is aimed at knowledge work and the other at developers. We spent real time with this before writing, which is why the practical notes below go beyond the official description.
Straight out of the box, Cowork reads and writes local files, splits work across parallel subagents, and produces real deliverables — spreadsheets with working formulas, slide decks, formatted documents. What it does not know is your CRM, your brand voice, or the way your team runs a quarterly review. That gap is exactly what Claude Cowork plugins are built to close.
Key takeaway
Claude Cowork plugins are packaged bundles that add skills, MCP connectors, subagents, slash commands, and hooks in a single install — sourced from a marketplace, your organization's Git repository, or an uploaded file, with each component individually toggleable; they work in Cowork and Claude Code but not in ordinary Chat.
What Claude Cowork plugins actually are
A plugin is one installable package that can add several capabilities to Cowork at once, instead of you wiring up a connector here and a skill there. The format came from Claude Code, so the same package works in both — but never in Chat.
A plugin is a package that extends what Claude can do in Cowork. Installing one can add several capabilities at once instead of making you wire them up one at a time. That single-step quality is the whole point: rather than adding a connector here, writing a skill there, and defining a subagent somewhere else, you install one package and Claude picks up all of it.
Plugins are supported in Cowork and in Claude Code. They are not used in Chat, so a plugin you install will not change how the ordinary conversational Claude behaves.
The idea started in Claude Code, where developers distribute plugins as versioned, shareable directories with a manifest describing their identity and contents. Cowork adopted the same model, which is why the file format is shared between the two and why a marketplace you build works for both. If you have already read about the Claude Code plugin marketplace, the mental model transfers almost perfectly.
What a plugin can contain
A manifest declares any mix of four component types: skills, MCP connectors, subagents, and hooks. Packages can also expose slash commands. After installing, open the package to see exactly which components ship inside, then toggle each one on or off.
A plugin's manifest declares any combination of four component types. Knowing which is which makes it much easier to judge what a plugin will actually do once it is running.
| Component | What it adds | Example |
|---|---|---|
| Skills | Reusable instructions that teach Claude a workflow | A brand-voice skill that activates when drafting external emails |
| Connectors | MCP servers that give Claude access to an external service | A CRM connector that reads and updates deal records |
| Agents | Specialized subagents Claude can delegate to | A subagent doing competitive research while another drafts a proposal |
| Hooks | Scripts that run at defined points in a session | A check that runs before Claude writes to a shared folder |
Plugins can also expose slash commands, which are the explicit, user-triggered entry points into a workflow. After installing, you open the plugin to see what it provides: skills and agents appear as tabs, while connectors and hooks get their own pages. That review screen matters — it is where a vague package description turns into a concrete list of what this thing can touch.

How to install Claude Cowork plugins
Open Customize → Plugins in the sidebar, select Browse plugins, pick a package and click Install, then review its components. Authentication prompts appear for any bundled connector. To sideload instead, use the upload option on the same page.
Installing is deliberately short. Open Customize in the sidebar, then Plugins, and work through three steps:
- Browse the marketplace. Select Browse plugins to see what is available. The default marketplace is Anthropic's official catalog, and you can add other marketplaces by URL.
- Install. Pick a plugin and click Install. If it bundles a connector that needs authentication, you get prompted to sign in to that service.
- Review components. Open the installed plugin and look at its skills, connectors, agents, and hooks. Enable or disable individual components as needed.
That third step is the one people skip and later regret. A plugin is not all-or-nothing — if a package ships five useful skills and one connector you would rather not authorize, turn that connector off and keep the rest.
To install from a file instead of a catalog, use the upload option on the Plugins page and select the plugin package. This is the path for a package a colleague handed you directly, or one you built yourself and have not published anywhere.
Anthropic has also open-sourced a set of plugins it uses internally, spanning areas such as sales, finance, legal, marketing, customer support, product management, and data work. There is even a Plugin Create plugin whose job is helping you build new ones from scratch — a reasonable starting point if you want to learn the format by example rather than by reading a manifest spec.
Using a Git repository as your own marketplace
Any Git repository holding plugin packages can act as a marketplace. Select Add marketplace, paste the repository URL or the owner-and-repo shorthand, and its packages appear beside every other catalog. GitHub Enterprise works, as do public GitLab and Bitbucket repositories.
Any Git repository containing plugin packages can act as a marketplace. This is the normal way a team distributes its own Claude Cowork plugins without publishing to the public catalog. GitHub is supported, including GitHub Enterprise, and public repositories on GitLab and Bitbucket work too.
On the Plugins page, select Add marketplace and enter the repository URL. Cowork accepts the standard full GitHub URL form as well as the shorter owner-and-repo shorthand. Plugins defined in that repository then appear alongside plugins from every other marketplace you have added, and install identically. Clicking Update on a marketplace pulls the latest plugins from its repository.
For teams, this is the quietly important feature. Your internal expense policy, your house style, your deal-desk process — all of that can live in one repository that everyone points Cowork at, and updates arrive by pulling rather than by asking people to reconfigure anything.
Limits worth knowing
Current defaults are 200 MB uncompressed per package, 5,000 files per package, a 512 MB marketplace repository archive, 500 packages per marketplace, and 25 marketplaces total. The in-app viewer previews files up to 1 MB, though Claude still reads larger ones at runtime.
Claude Cowork plugins are generous on size but not unbounded. These are the current defaults.
| Limit | Value |
|---|---|
| Plugin package size (uncompressed) | 200 MB |
| Files per plugin package | 5,000 |
| Marketplace repository archive | 512 MB |
| Plugins per marketplace | 500 |
| Marketplaces you can add | 25 |
The in-app skill viewer previews individual files up to 1 MB. Larger ones show as too large to preview in the file list, but Claude can still use them at runtime — the limit is about the viewer, not the model.
How plugins relate to connectors and skills
Packages do not replace connectors or skills — they wrap them. A lone connector or a loose skill is fine when you are setting something up for yourself. Bundling only matters when several colleagues need the identical setup installed in one click.
This trips people up, so it is worth stating plainly: plugins do not replace connectors or skills. They package them.
A connector is an MCP server that gives Claude access to an external tool or data source, and you can add one on its own — our overview of Claude connectors walks through how that works independently of any plugin. A skill is a set of reusable instructions that teaches Claude a workflow, and the same is true there; Claude skills stand alone perfectly well. A plugin is the wrapper that lets you ship a connector, a few skills, a subagent, and a hook as one unit that a teammate installs in a single click.
So the decision is less about capability than about distribution. Setting something up for yourself only? A bare connector or a loose skill is fine. Setting something up that five colleagues need identically? That is what Claude Cowork plugins are for.

Safety notes you should actually read
Anthropic curates its official catalog and screens community submissions, but says plainly that it cannot control which MCP servers or files a package includes. Install only from developers you trust, read the component list before enabling, and keep file access narrowly scoped.
An agent with more tools and more file access is more useful and more consequential. Both things are true at once, and Claude Cowork plugins sit right at that intersection.
Anthropic curates the official marketplace, and inclusion is at its discretion. A community marketplace exists whose plugins have passed automated validation and safety screening, and plugins carrying an Anthropic Verified badge have been through additional quality and safety review. That is real, but Anthropic is also direct about the ceiling: it does not control what MCP servers, files, or other software a plugin includes, and it cannot verify that every plugin works as intended. Community plugins may install unverified third-party software. The official guidance is blunt and correct — only install plugins from developers you trust.
Some practical habits worth adopting:
- Read the component list before you enable anything, not after. A plugin that quietly bundles a hook is doing something on a schedule you did not explicitly ask for.
- Scope file access to the folders a task genuinely needs, rather than pointing Cowork at your whole home directory and hoping.
- Treat an unfamiliar third-party connector the way you would treat an unfamiliar browser extension asking to read every page.
- Prefer a repository you or your company controls over a random public one when the plugin will touch anything sensitive.
Cowork currently runs Claude inside an isolated virtual machine environment, and the newest models — Claude Opus 5 among them — are noticeably better at flagging when a step looks out of scope. Neither of those is a substitute for reviewing what you install.
Because we run this site's whole workflow in Claude Code, we work with the same skills, MCP connectors, subagents, and hooks that Claude Cowork plugins package up — just wired in by hand rather than bundled. Living with that setup daily is what convinced us the review-the-component-list step is not optional busywork: the parts that bite quietly are the hooks, because a hook runs on its own without you asking it to each time. Our standing rule now is that anything carrying a hook or a connector gets read line by line before it is enabled, and file access stays scoped to the one folder a job actually needs.
Organization-managed plugins
On Team and Enterprise plans, administrators can mark packages as required. Those install automatically, display a note saying your organization requires them, and cannot be uninstalled from your side — only an administrator can remove them.
On Team and Enterprise plans, administrators can require certain Claude Cowork plugins for everyone in the organization. Required plugins install automatically and display a note saying the plugin is required by your organization, and you cannot remove them. Only an administrator can. If you are trying to work out which plans get what, our rundown of who gets Claude Cowork is the shorter answer.
Plugin support in Cowork is currently in beta for all paid Claude users, with plugins saved locally to your machine and broader org-wide sharing and management on the roadmap.
Updating and removing plugins
Cowork checks each marketplace for newer versions and warns you before an update would overwrite files you edited locally. Uninstall anything you installed yourself from Customize → Plugins. Organization-required packages stay in place until an administrator removes them.
Cowork checks for updates from whichever marketplace a plugin came from. If you have edited a plugin's files locally, it detects the change and warns you before an update would overwrite your version — a small touch that saves real frustration for anyone who has customized a package in place.
To remove a plugin you installed yourself, open it under Customize → Plugins and click Uninstall. Organization-managed plugins can only be removed by an administrator, which is the intended behavior rather than a bug.
For the authoritative version of all of this, go to the Cowork plugins guide and the broader plugins overview. Product surfaces move, and the docs are the source of truth when a detail here and a detail there disagree.
Frequently Asked Questions

Written by
Edith
Writing about Claude and the Anthropic toolkit — models, Claude Code, pricing, features, and fixes, in clear, practical, hands-on guides tested by daily use.
View all posts →


