Community Ecosystem

Module 15.4: Community Ecosystem

Estimated time: ~30 minutes

Prerequisite: Module 15.3 (Claude Code Skills)

Outcome: After this module, you will be able to add a plugin marketplace, inspect a plugin’s hooks and MCP config before installing it, install it at project scope so your team gets it, and restrict which marketplaces and MCP servers a team may use.

1. WHY — Why This Matters

Someone on the team pastes “install this plugin, it’s great” with a GitHub link. It ships a SessionStart hook, a Stop hook and an MCP server that wants a token. Everyone installs it because it is popular; nobody reads it. Six weeks later those hooks still run at every session start and every turn end in a repo that holds bank credentials.

The docs are blunt: “Plugins and marketplaces are highly trusted components that can execute arbitrary code on your machine with your user privileges.” So: read before installing, as a team rule.

2. CONCEPT — Core Ideas

Two steps: marketplace, then plugin

A marketplace is a git repo (or directory, or URL) with .claude-plugin/marketplace.json listing plugins. Add it once, then install plugins from it as name@marketplace.

graph LR
    M["marketplace<br/>.claude-plugin/marketplace.json"] -->|"/plugin install name@marketplace"| P["plugin<br/>.claude-plugin/plugin.json"]
    P --> S["skills/ · agents/"]
    P --> H["hooks/hooks.json<br/>runs on events"]
    P --> C[".mcp.json<br/>servers + tokens"]
    S & H & C --> U["all run as you"]

Where plugins come from

Source	Add / install	What is there
`claude-plugins-official`	Auto-registered; browse in `/plugin` → Discover or claude.com/plugins	Anthropic’s catalogue, e.g. `/plugin install github@claude-plugins-official`
`anthropics/claude-plugins-community`	`/plugin marketplace add anthropics/claude-plugins-community` → `name@claude-community`	Community plugins
`anthropics/claude-code`	`/plugin marketplace add anthropics/claude-code` → `claude-code-plugins`	Anthropic’s own repo: `commit-commands`, `security-guidance`, `plugin-dev`, `hookify`, …
`anthropics/skills`	`/plugin marketplace add anthropics/skills` → `anthropic-agent-skills`	`document-skills` (docx/pdf/pptx/xlsx), `example-skills`, `claude-api`, … many Apache 2.0
Your own repo	`/plugin marketplace add your-org/claude-plugins`	Internal plugins
Curated lists	e.g. awesome-claude-code	A list, not a review

The docs’ warning covers all of them: “Anthropic doesn’t control what MCP servers, files, or other software are included in plugins and can’t verify that they work as intended.”

Scopes: who gets the plugin

Scope	Written to	Use for
`user` (CLI default)	`~/.claude/settings.json`	Your machine only
`project`	`.claude/settings.json`: `extraKnownMarketplaces` + `enabledPlugins`	Commit it; teammates get it after folder trust
`local`	`.claude/settings.local.json`	This checkout only
`managed`	Managed settings	Org-wide, not overridable

The pre-install checklist

Before /plugin install, open the source and answer five questions:

Who publishes it? Org, commit history, known marketplace.
hooks/hooks.json: which events (UserPromptSubmit and Stop fire every turn, SessionStart “when a session begins or resumes”) and what the scripts do.
.mcp.json: which endpoints, which tokens go into headers or env.
Skills: allowed-tools, !`commands`, disable-model-invocation on side effects.
Does it need all that? A commit helper does not need a Stop hook.

Team controls (enforced, not advisory)

Setting	Scope	Effect
`enabledPlugins`	any; managed `false` blocks every scope	On/off per plugin; project overrides user
`extraKnownMarketplaces`	any; repo settings honored after trust	Auto-add marketplaces for a project
`strictKnownMarketplaces`	managed only	Allowlist of sources; `[]` locks down everything, official included
`blockedMarketplaces`	managed	Denylist
`allowedMcpServers`	any; “Deploy it in managed settings to enforce it”	Allowlist by `serverName`, `serverCommand`, `serverUrl`; covers plugin servers

3. DEMO — Step by Step

Run in ~/cc-lab. The steps use the claude plugin … shell form so output is reproducible; /plugin … inside a session does the same.

Step 1: See which marketplaces you already have

# docs: plugin-marketplaces
claude plugin marketplace list

# Output may vary
Configured marketplaces:

  ❯ claude-plugins-official
    Source: GitHub (anthropics/claude-plugins-official)
  …

Step 2: Add Anthropic’s repo marketplace at project scope

# docs: plugin-marketplaces, discover-plugins
claude plugin marketplace add anthropics/claude-code --scope project
cat .claude/settings.json

# Output may vary
Adding marketplace…Cloning via SSH: git@github.com:anthropics/claude-code.git
Refreshing marketplace cache (timeout: 120s)…
Clone complete, validating marketplace…
✔ Successfully added marketplace: claude-code-plugins (declared in project settings)
{
  "extraKnownMarketplaces": {
    "claude-code-plugins": {
      "source": {
        "source": "github",
        "repo": "anthropics/claude-code"
      }
    }
  }
}

The name comes from the marketplace’s marketplace.json, not the repo.

Step 3: Read before you install

The marketplace is a public repo, so read what it will run:

# docs: plugins-reference (hooks/hooks.json, .mcp.json layout)
curl -s https://raw.githubusercontent.com/anthropics/claude-code/main/plugins/security-guidance/hooks/hooks.json \
  | jq -c '.hooks | keys'
for f in commit commit-push-pr clean_gone; do
  echo "== $f.md"
  curl -s https://raw.githubusercontent.com/anthropics/claude-code/main/plugins/commit-commands/commands/$f.md \
    | grep allowed-tools
done
curl -s https://raw.githubusercontent.com/anthropics/claude-plugins-official/main/external_plugins/github/.mcp.json

# Output may vary
["PostToolUse","SessionStart","Stop","UserPromptSubmit"]
== commit.md
allowed-tools: Bash(git add:*), Bash(git status:*), Bash(git commit:*)
== commit-push-pr.md
allowed-tools: Bash(git checkout --branch:*), Bash(git add:*), Bash(git status:*), Bash(git push:*), Bash(git commit:*), Bash(gh pr create:*)
== clean_gone.md
{
  "github": {
    "type": "http",
    "url": "https://api.githubcopilot.com/mcp/",
    "headers": {
      "Authorization": "Bearer ${GITHUB_PERSONAL_ACCESS_TOKEN}"
    }
  }
}

security-guidance hooks four events, including every prompt and every stop: that is its job, but know it before it runs on a repo with secrets. commit-commands is three command files: /commit can only add, status and commit; /commit-push-pr may also push, create branches and open PRs; and /clean_gone, which deletes local branches and worktrees, pre-approves nothing, so every command prompts. The official github plugin is an HTTP MCP server that sends your GITHUB_PERSONAL_ACCESS_TOKEN to api.githubcopilot.com. None of this is hidden; it is only unread.

Step 4: Install at project scope

# docs: discover-plugins, plugins-reference
claude plugin install commit-commands@claude-code-plugins --scope project
cat .claude/settings.json

# Output may vary
Installing plugin "commit-commands@claude-code-plugins"...✔ Successfully installed plugin: commit-commands@claude-code-plugins (scope: project)
{
  "extraKnownMarketplaces": { "claude-code-plugins": { … } },
  "enabledPlugins": {
    "commit-commands@claude-code-plugins": true
  }
}

Step 5: List it from inside a session

Start claude and type /plugin list:

# Output may vary
❯ /plugin list
  ⎿  Installed plugins:
       • commit-commands@claude-code-plugins (v1.0.0, project) ✔ enabled
       …

Step 6: Remove it cleanly

# docs: plugins-reference, plugin-marketplaces
claude plugin uninstall commit-commands@claude-code-plugins --scope project
claude plugin marketplace remove claude-code-plugins

# Output may vary
✔ Successfully uninstalled plugin: commit-commands (scope: project)
✔ Successfully removed marketplace: claude-code-plugins

Removing a marketplace “also uninstalls any plugins you installed from it”.

4. PRACTICE — Try It Yourself

Exercise 1: Audit a plugin you did not write

Goal: Apply the checklist to anthropics/skills.

Instructions:

claude plugin marketplace add anthropics/skills --scope local.
Read the repo’s .claude-plugin/marketplace.json and list the plugin names.
For document-skills, note the tools each SKILL.md needs and whether any hook or .mcp.json exists. Then claude plugin marketplace remove anthropic-agent-skills.

Expected result: One sentence saying what the plugin runs and with which tools.

✅ Solution

curl -s https://raw.githubusercontent.com/anthropics/skills/main/.claude-plugin/marketplace.json \
  | jq -c '.plugins[] | {name, skills, hooks, mcpServers}'

# Output may vary
{"name":"document-skills","skills":["./skills/xlsx","./skills/docx","./skills/pptx","./skills/pdf"],"hooks":null,"mcpServers":null}
{"name":"example-skills","skills":["./skills/algorithmic-art",…],"hooks":null,"mcpServers":null}
{"name":"claude-api","skills":["./skills/claude-api"],"hooks":null,"mcpServers":null}
{"name":"academy-guide","skills":["./skills/academy-guide"],"hooks":null,"mcpServers":null}
{"name":"discernment-nudge","skills":["./skills/discernment-nudge"],"hooks":null,"mcpServers":null}

document-skills: four skills with scripts/, no hooks, no MCP servers, no allowed-tools, so each script goes through your normal permission flow. Their SKILL.md files say license: Proprietary; the README’s Apache 2.0 covers other skills.

Exercise 2: Lock a team down

Goal: Managed settings that allow only the official marketplace, your org’s repo, and one MCP server.

✅ Solution

{
  "strictKnownMarketplaces": [
    { "source": "github", "repo": "anthropics/claude-plugins-official" },
    { "source": "github", "repo": "your-org/claude-plugins" }
  ],
  "allowedMcpServers": [
    { "serverUrl": "https://mcp.internal.example.com/*" }
  ]
}

strictKnownMarketplaces is managed-only and gates marketplace sources. allowedMcpServers works in any settings file but is enforced org-wide only from managed settings, and it does cover servers a plugin ships.

Exercise 3: Publish the plugin from Module 15.5 internally

Goal: A private marketplace your team adds with one command.

✅ Solution

{
  "name": "acme-tools",
  "owner": { "name": "ACME Platform Team" },
  "plugins": [
    {
      "name": "cc-lab-plugin",
      "source": "./plugins/cc-lab-plugin",
      "description": "Test-writing skill plus post-write test hook"
    }
  ]
}

Save as .claude-plugin/marketplace.json next to plugins/cc-lab-plugin/, run claude plugin validate . (✔ Validation passed with warnings until you add a description), push, then /plugin marketplace add your-org/acme-tools and /plugin install cc-lab-plugin@acme-tools.

5. CHEAT SHEET

Command / setting	Purpose
`/plugin`	Menu: Discover · Installed · Marketplaces · Errors · Stats
`/plugin marketplace add owner/repo` (`./dir`, git URL, `owner/repo@ref`)	Register a marketplace
`/plugin marketplace list` · `update <name>` · `remove <name>`	Manage marketplaces
`/plugin install name@marketplace`	Install; prompts for scope
`/plugin install name --marketplace owner/repo`	Add + install in one step (v2.1.275+)
`/plugin uninstall name@marketplace` · `enable` · `disable` · `list`	Manage plugins
`claude plugin install name@marketplace -s project`	Shell form; user scope unless `-s`
`claude plugin validate .`	Check a `marketplace.json` or `plugin.json`
`/reload-plugins`	Activate without restarting
`"enabledPlugins": {"name@market": true}`	Per-scope on/off; managed `false` blocks everywhere
`"extraKnownMarketplaces"`	Auto-add for a project (after folder trust)
`"strictKnownMarketplaces": []`	Managed lockdown; list sources to allow
`"allowedMcpServers"`	MCP allowlist; enforced from managed settings

6. PITFALLS — Common Mistakes

❌ Mistake	✅ Correct Approach
Install counts, star counts or awesome-lists as due diligence	They say something exists and is popular. Run the five-question checklist yourself
User scope for a team tool	`--scope project` writes `enabledPlugins` to `.claude/settings.json`; commit it
”It’s an official plugin, so it’s safe”	Same warning applies. `github@claude-plugins-official` sends a token to an HTTP server: fine if you meant that
Putting `strictKnownMarketplaces` in `.claude/settings.json`	It is managed-only. Project settings get `enabledPlugins` / `extraKnownMarketplaces`
Treating a plugin’s hook as harmless because it “only reminds”	It runs a script with your privileges on each event; read the script, not the description

7. REAL CASE — Production Story

Scenario: A Hanoi outsourcing company runs Claude Code across a dozen client repos, some with banking credentials in CI. Developers added marketplaces freely; nobody could say which hooks ran where.

Problem: A client security review asked “what third-party code executes when your engineers open our repo?” The honest answer was “we don’t know”.

Solution: Plugins became a reviewed artefact: every proposal is a PR to your-org/claude-plugins copying the plugin in with a filled checklist (events in hooks/hooks.json, endpoints and tokens in .mcp.json, allowed-tools per skill). Managed settings set strictKnownMarketplaces to the official marketplace plus that repo, and allowedMcpServers to two internal servers. Each project’s .claude/settings.json carries extraKnownMarketplaces and enabledPlugins, so a checkout declares what runs in it. Same rule Anthropic applies to its agents (S4): “Give every agent a single-purpose identity with the minimum permissions for its job”.

Result: The client question now has a file as its answer; a new hook cannot reach a repo without a PR review. Adding a plugin takes a day, not a minute, which is the point.

Next: Module 15.5: Custom Skill Development →