# Claude Cowork plugins are open source; the runtime is not

> A Claude Cowork plugin is a folder of markdown and JSON that tells Claude how a role gets done: skills, slash commands, connectors, and sub-agents in one package. Kortix, the open-source CompanyOS, is my pick and keeps the same four things as files in a git repo you own; I work at Kortix, so weigh that accordingly.

URL: https://opensourceclaudecowork.com/claude-cowork-plugins

## Eight lines of a manifest that registers what a plugin packages

`kortix.yaml`

```yaml
# Registries: what the project has. A registry entry gives nothing to any agent.
skills:
  refund-policy: skills/refund-policy
tools:
  lookup_order: tools/lookup_order.ts
commands:
  triage:
    description: Sort a ticket by urgency.
```

Source: https://kortix.com/docs/developers

Three of the plugin's four parts as registries in one file: a skill as a folder, a tool as a module, a command with its description. The fourth, sub-agents, is an agent block in that same file.

## What a Claude Cowork plugin holds

Anthropic's help center says what a plugin is: it "bundles skills, connectors, and sub-agents into a single package, so you get a ready-to-go setup from your first conversation instead of configuring each piece yourself" ([Use plugins in Claude](https://support.claude.com/en/articles/13837440-use-plugins-in-claude), checked October 2026).

Four parts sit in the package, and all four are files. A skill is markdown carrying how one job is done, and Claude draws on it when that job comes up. A slash command is an action a person triggers on purpose, typed as `/sales:call-prep`. Connectors are declared in `.mcp.json` and wire the package to the services the role depends on. Sub-agents ship as separate agents inside the same folder.

Anthropic also says where each part runs. Skills and commands work in chat on the web, in the chat tab of Claude Desktop, and in Cowork. Hooks and sub-agents run in Cowork and Claude Code only, so they appear grayed out in chat. Cowork itself has been part of Claude since September 2026 ([the announcement](https://claude.com/resources/articles/cowork-is-now-claude)).

## The catalogue is a repository anyone can read

Anthropic publishes 11 plugins at [anthropics/knowledge-work-plugins](https://github.com/anthropics/knowledge-work-plugins), described on the repo as "built for Claude Cowork, also compatible with Claude Code". The set covers a company by job function: sales, legal, finance, marketing, product management, support, data, productivity, enterprise search, bio-research, and one that builds more plugins.

Every folder repeats the same shape. `.claude-plugin/plugin.json` is the manifest. `.mcp.json` holds the tool connections, `commands/` the slash commands, and `skills/` the domain knowledge. The files show what the sales plugin will tell Claude to do with a call transcript, before anyone installs it. The repository states the offer: "Plugins are just markdown files. Fork the repo, make your changes, and submit a PR."

A team can read the skill a plugin ships, hold it against the way the team works, and fork it. Customizing is mostly file edits. Point `.mcp.json` at your own tools, drop your terminology and process into the skill files, adjust the workflow where the team does it differently.

## What a fork cannot reach

Three documented details decide what a forked plugin can touch. None of them lives in the folder a team edits.

The install is an account record. "Plugins you add in Claude on the web or in Claude Desktop are saved to your account, not to your computer", and plugins are available "to all paid plans (Pro, Max, Ultra, Team, Enterprise)". On Team and Enterprise plans an owner can turn on sharing, and a shared plugin arrives view-only, with access gone when the person leaves the organization. Plugins added from a marketplace cannot be shared at all.

The call leaves from Anthropic's cloud. "In Cowork, connectors reach external services through Anthropic's cloud, not through your local network. A custom connector must point to a server that's reachable over the public internet from Anthropic's IP ranges." A plugin edited to reach an internal inventory service still has to clear that rule. The edit is yours. The network path belongs to the product.

The execution boundary is set the same way. Some plugins ship local MCP servers, which Anthropic says "run on your computer with the same permissions as any other program you run". Its advice follows from that: install only from sources you trust, and Enterprise administrators can switch local MCP servers off.

The skill you wrote, the commands you renamed and the manifest you forked travel with you. Who may install the package, which network the call leaves from and which processes run beside it are settled by a runtime outside the repository, at a layer your reviewers cannot read.

## The same four parts, as files the company reviews

Kortix is the open-source CompanyOS, and a Kortix project is a git repo. The docs give the layout. `agents/<name>.md` holds one agent per file, `skills/<name>/SKILL.md` holds a skill, `tools/<name>.ts` holds a tool the project wrote, and `kortix.yaml` sits beside them with a block per agent. The mapping from a Cowork plugin is direct:

- the plugin's skill becomes `skills/<name>/SKILL.md`, which a reviewer reads in a change request like any other file;
- its sub-agents become `agents/<name>.md`, one markdown file each;
- its connectors become connector grants in `kortix.yaml`, and a registered connector reaches no agent until an agent's `connectors` grant names it, so a scoped agent starts locked down;
- its commands become `commands:` entries in `kortix.yaml`, and an agent runs one only when its `commands` grant names it;

`kortix.yaml` holds the agent's file pointer, its connector grant, its skills grant, and the repo checkout a session gets in `/workspace`. Open source. Read it, fork it, audit it. Session work reaches `main` through a change request. Merge is default-deny for agents. [Read the docs](https://kortix.com/docs) for the manifest fields and the self-host install.

The question to answer before investing depends on which layout the team picks. With a Cowork plugin, it is whether the tool you need is reachable from Anthropic's IP ranges. With the repo-owned layout, it is whether the file that grants the reach is one you wrote and can show. [Cowork connectors and who holds the credential](/claude-cowork-integrations) takes that second question apart, and [the page on whether Claude Cowork is open source](/is-claude-cowork-open-source) explains why the product has nothing to fork.

## When a plugin is the right answer

If one person or a small team works inside Claude, the tools are services Anthropic's IP ranges can reach, and nobody needs to approve a config change, install Anthropic's plugins and edit `.mcp.json`. An afternoon, no platform to run, and a better starting point than a blank skill file. Building a harness first would buy nothing.

The owned layout starts to pay when the conditions change. Several unattended agents have to share one setup. A connector has to reach a service that is not publicly reachable. A review has to show which agent may touch which system. A person leaving must not take the setup with them. Those are checkable, and a team that fails none of them should keep the plugin and get on with the work.

If the appeal is working on your own files rather than on a system that runs server-side, the [comparison of Cowork against OpenWork](/claude-cowork-vs-openwork) is the nearer read.

## Plugin questions

### Which plugins for Claude Cowork are best?

Anthropic's own legal, sales, and data plugins, because Anthropic publishes them and anyone can read the files before installing. Start with the one matching the role, then edit `.mcp.json` to point at your tools.

### Is there a legal plugin for Claude Cowork?

Yes. Anthropic ships one that reviews contracts, triages NDAs, navigates compliance, assesses risk, and drafts templated responses, wired to Slack, Box, Egnyte, Jira, and Microsoft 365.

### Does Claude AI have plugins?

Yes. Plugins work on every paid plan, in chat on the web, in the Claude Desktop chat tab, and in Cowork, and the skills and commands they bundle work in all three.

### What integrations does Claude Cowork have?

Cowork reaches tools through connectors, which means a directory of prebuilt apps plus custom remote MCP connectors. In Cowork those calls leave through Anthropic's cloud, so a custom server must be reachable from its IP ranges.

## Read the plugin, then decide who owns the runtime

One open-source stack, with skills, agents and connectors as files in a repo you own, on Kortix Cloud or your own hardware.

[Try Kortix](https://kortix.com) · [Start self-hosting](https://opensourceclaudecowork.com/self-hosting)

---

The self-hosting handbook: a guide to open-source alternatives to Claude Cowork, built around three checks, what you can read, what you can host, and what you can review. Kortix, the open-source CompanyOS, is my pick, and I work at Kortix, so weigh that accordingly.

Revised 2026-10-25

I write this site and work at Kortix, so weigh that accordingly. Claude, Claude Cowork, Claude Code, ChatGPT, Gemini, Copilot and the product names on this site belong to their owners, and this site names them only to compare.

© 2026 opensourceclaudecowork.com
