← Default view
/p Claude Code Plugin Marketplace /plugin › Discover
Discover Installed Marketplaces Errors
Talk-prep brief · Session 3 of a series

Plugins & the Marketplace

What plugins actually are, how the official + community marketplaces work, the few that earn their keep, and the trust footgun to flag in any 1–2 hour deep-dive.

survey · 10 min read 21 sources cited updated 2026-05-23 Read default view →
Tier Official · auto-shipped Community · auto-screened Third-party · caveat emptor /plugin marketplace add <name> Sorted by: editorial → stars
Verdict
for the talk

A Claude Code plugin is a packaging unit — a directory bundling any of skills, agents, slash commands, hooks, MCP servers, LSP servers — installed via /plugin from a marketplace (a marketplace.json catalog hosted on GitHub or anywhere reachable). The official Anthropic marketplace ships automatically (~101 plugins, 33 Anthropic + 68 partner as of May 2026)[20]; the community marketplace is one /plugin marketplace add away, with submissions gated by automated safety screening[7]. Spend ~25 min on anatomy and install flow, ~15 min on building, and ~15 min on the trust boundary — plugins execute arbitrary code with the user's privileges, and there is already a CVE history to point to[13].

Editor's pick · live-demo material

Anthropic's reference example. Installs cleanly, narrates well, and gives the audience an immediate "oh, that's all it is?" moment[2].

101plugins
Official marketplace[20]
2,500+
3rd-party marketplaces indexed[14]
6,700+
Skills catalogued community-wide[14]
4/10
Tested plugins judged worth keeping[11]

What's inside a plugin

Identified by .claude-plugin/plugin.json; component dirs go at the plugin root, NOT inside .claude-plugin/ — common landmine[3].

Skills
skills/<name>/SKILL.md
Model-invoked capability with a YAML description Claude reads to decide when to use it[3].
Slash commands
commands/
User-invoked /plugin-name:foo workflows[9].
Agents
agents/
Specialised sub-agent definitions invoked by the host or by other agents[3].
Hooks
hooks/hooks.json
Event handlers — SessionStart, PreToolUse, Stop[9].
MCP servers
.mcp.json
Pre-configured external tool integrations bundled with the plugin[9].
❮>
LSP servers
.lsp.json
Language-server bindings for real-time code intelligence after every edit[3].
Background monitors
monitors/monitors.json
tail -F-style watchers that push notifications into the session[3].
Default settings
settings.json
Plugin-scoped defaults; only agent + subagentStatusLine honoured today[3].

Plugins worth a slide

Don't recommend the whole shelf — pick the ones that earn their keep. Drawn from the official marketplace's dev workflows[2] and one of the few honest reviews[11].

/review:bugs
/review:compliance
/review:context
/review:… (5 agents)
Official 5 agents
by Anthropic · claude-plugins-official

Five parallel review agents — compliance, bugs, context, performance, and one more — running concurrent passes over the diff[2].

Concrete multi-agent demo without leaving the CLI.
/plugin install pr-review-toolkit@claude-plugins-official
/plugin-dev:create-plugin
/plugin-dev:agent-creator
/plugin-dev:skill-reviewer
/plugin-dev:validator
Official meta
by Anthropic · claude-plugins-official

The meta-plugin. Bundles a scaffolder, agent-creator, skill-reviewer, and validator so you can build a plugin on stage[9].

Use it to live-build a plugin during the talk.
/plugin install plugin-dev@claude-plugins-official
! PreToolUse hook
eval / unsafe deserialize
SQL/cmd injection · XSS
⚠ blocked · review · allow
Official hook callback ↔ AI-Sec
by Anthropic · claude-plugins-official

PreToolUse hook that warns on injection, XSS, eval, and unsafe deserialize patterns before the tool call fires[9].

Direct callback to session 2 (AI Security).
/plugin install security-guidance@claude-plugins-official
# bridge to your audience's
# own automation work
/agent-sdk-dev:scaffold
/agent-sdk-dev:tools-doc
Official SDK
by Anthropic · claude-plugins-official

Skills and agents for building with the Claude Agent SDK — scaffolders, tool-doc generators, eval harnesses[2].

Bridge into the audience's own automation work.
/plugin install agent-sdk-dev@claude-plugins-official
// after every edit
typescript-lsp
pyright-lsp   
rust-analyzer
→ diagnostics + go-to-def
Official single highest-leverage
by Anthropic · claude-plugins-official

Any one LSP pack gives Claude diagnostics and go-to-def after every edit. IDE-grade feedback, in-terminal[2].

Single highest-leverage install for code work.
/plugin install typescript-lsp@claude-plugins-official
wshobson/agents GitHub preview
Third-party inspect first 80+ agents
36k
by Seth Hobson · github

The headline community pack — 80+ specialised sub-agents, each scoped to a single concern (security review, refactor, perf, etc.)[15].

Show how to inspect a third-party set before installing.
/plugin marketplace add wshobson/agents
davila7/claude-code-templates GitHub preview
Third-party CLI · templates
by Dan Avila · github

Templating CLI for plugins, agents, and commands — useful as a survey of what people are actually building, not just downloading[16].

Audit the shape of the community before recommending anything.
npx claude-code-templates
obra/superpowers GitHub preview
Third-party heavy skill set most-starred
204k
by Jesse Vincent (obra) · github

Among the most-installed third-party skill sets in the ecosystem — a curated grab-bag of SKILL.md files, agents, and slash commands[17].

Reference for "what does scale look like on this surface?"
/plugin marketplace add obra/superpowers
claudemarketplaces.com preview
Third-party directory community index
170k visitors/mo
community-run discovery directory

Indexes 6,700+ skills, 2,500+ marketplaces, and 840+ MCP servers — the de-facto external catalog when /plugin isn't enough[14].

Audience research and "what's out there" homework.
browse ↔ https://claudemarketplaces.com
Honest aside. Of 10 popular plugins tested in early 2026, only 4 were judged worth keeping — and the failures were not subtle. Treat the marketplace like npm, not the App Store[11].

Partner integrations in the official marketplace

Worth name-dropping when you introduce the official tier. Anthropic-vetted, but still third-party code[6].

GitHub
GitLab
Atlassian
Linear
Asana
Notion
Figma
Vercel
Firebase
Supabase
Slack
Sentry
Stripe
Playwright
Auth0
Datadog
Cloudflare
aws-* family

The three marketplace tiers

One schema, three trust contracts. The schema is published at schemastore.org/claude-code-marketplace.json[4].

auto on first launch26kAnthropic-curated

33 Anthropic-built + 68 partner integrations, vetted at Anthropic's discretion — no public application process[20].

# ships automatically

claude-plugins-community

opensubmission
manual add119auto-screened

Anyone can submit via claude.ai/settings/plugins/submit. Anthropic runs automated validation and safety screening; each entry pins to a commit SHA[7].

/plugin marketplace add anthropics/claude-plugins-community
manual addgithub / git / URL / pathcaveat emptor

Any GitHub repo, git URL, HTTPS URL, or local path. No curation, no automated screening. Treat exactly like an npm install[14].

/plugin marketplace add <owner/repo>

Install scopes

Maps cleanly to the standard Claude Code settings tiers. Pick the right one for the demo scenario[2].

user
~/.claude/settings.json
Personal toolkit across all projects. The default scope.
project
.claude/settings.json · committed
Team-shared via version control. Wire extraKnownMarketplaces here so teammates get prompted to install on first folder-trust[2].
local
.claude/settings.local.json · gitignored
"Just me, just here." Doesn't follow the repo or your other machines.
Security advisories · the bit that earns the slot

The trust boundary

A plugin executes arbitrary code on the user's machine with their user privileges, just like an npm package[2]. Anthropic's own docs say so in a <Warning> box. The CVE history below is why session 3 follows session 2.

CVSS 8.7 Oct 2025 · patched
CVE-2025-59536

RCE via hooks / MCP config in .claude/settings.json, before the trust dialog appeared.

Check Point → research report [13]
CVSS 5.3 Jan 2026 · patched
CVE-2026-21852

API key exfiltration via ANTHROPIC_BASE_URL override — a redirect-the-endpoint attack.

Check Point → research report [13]
convention Apr 2026
"TrustFall"

Cloning + opening a hostile repo could trigger code execution before any explicit user consent. A whole-class attack, not a single bug.

Dark Reading → TrustFall write-up [19]
supply-chain Q1 2026
Marketplace dependency hijack

Malicious plugin silently reroutes pip install to an attacker-controlled mirror; behaviour persists across sessions.

SentinelOne → PoC + writeup [12]
Defensive posture · recommend on a slide

The five-line checklist

  • Treat plugins like dependencies — review source, pin versions, prefer official + community over arbitrary GitHub URLs[12].
  • Turn on managed marketplace restrictions for orgs — admin-controlled allowlist, not user discretion[2].
  • v2.1.145 lists exactly what a plugin will add before you confirm — read the "Will install" panel[21].
  • Auto-update is on by default for official, off for third-party. Don't blindly flip it on for community sources[2].
  • Disable auto-update with DISABLE_AUTOUPDATER=1; plugin-only updates via FORCE_AUTOUPDATE_PLUGINS=1[2].

Suggested session shape · 60–90 min

Block-by-block run-of-show. Plugin-dev gets the live-build slot; trust boundary gets the closer.

10 min
Why plugins · plugin ≠ skill
Open /plugin, walk the four tabs.
15 min
Anatomy + manifest
Inspect commit-commands source in the official repo.
10 min
Install flow + scopes
Install commit-commands at user scope, live.
15 min
Build one live
mkdir → plugin.json → SKILL.md → --plugin-dir.
10 min
Distributing
Real-world marketplace.json + extraKnownMarketplaces.
15 min
Trust boundary + CVEs
Walk SentinelOne's hijack; demo --plugin-dir sandboxing.
15 min
Q&A / picks
Have wshobson/agents and plugin-dev ready to install on request.
Third in a series
Extending Claude Code: MCP, Skills, Plugins — a session-3 blueprint →