Best Claude Code Plugins (2026): 10 Ranked by an Operator Whose Production Agents Run Zero

If you came here for a straight list of the best Claude Code plugins, jump to the ranked ten. Every one was checked against the official marketplace on 3 November 2026, along with what it actually installs on your machine. Then read why the container that published this post has zero plugins installed, because that’s the part most lists leave out. Updated 4 November 2026.

Ask ten people what Claude Code plugins are and you’ll get ten different answers — a marketplace thing, a VS Code extension, some kind of app store. None of them quite right. Here’s the operator’s version, no hand-waving: a plugin is a bundle. It packs skills, slash commands, hooks, and MCP connections into one installable folder so you can add a whole capability to your agent with a single command instead of wiring four things by hand.

I run more than ten autonomous businesses out of Claude Code containers. That fleet runs on skills and hooks I wrote myself, and I check every plugin against one question: would I let it near a container holding live business credentials? But most of the “top plugins” content out there is either official docs that only teach you how to install, a paywalled review you can’t read, or a Reddit list with zero judgment. This is the free, honest, tested-in-production take: what plugins actually are, which handful earn their place on a solo operator’s machine, which ones are bloat, and how to build your own from a task you already do every week.

What Claude Code Plugins Really Are

Strip away the marketing and a plugin is four familiar things in one wrapper. If you’ve spent any time in Claude Code, you already know the ingredients:

• Skills — packaged instructions and knowledge the agent loads on demand (think “how to write our blog posts” or “how to reconcile invoices”).
• Slash commands — the /thing shortcuts you type to trigger a workflow.
• Hooks — scripts that fire automatically at set moments, like before a file save or after a tool runs. Hooks are the difference between an assistant that suggests and an agent you can trust to act. I went deep on this in my guide to Claude Code hooks and how operators turn a coding assistant into an agent you can trust.
• MCP servers — connections to outside tools and data, from your database to Notion to Stripe.

A plugin is simply those pieces zipped into a self-contained directory with a small manifest file (.claude-plugin/plugin.json) that gives it a name, a version, and an author. That manifest is what turns a loose pile of config into something you can version, share, and install cleanly.

Here’s the distinction that matters for a solo operator. Claude Code gives you two ways to add custom behavior. The first is standalone config: you drop files straight into your project’s .claude/ folder. Fast, personal, perfect for experiments. The second is a plugin: the same capability, but packaged so it travels. When your skill lives in a plugin, its commands get namespaced — you call /my-plugin:publish instead of a bare /publish — which means no collisions when you install a dozen of them. The official guidance is exactly what I’d tell you: start standalone while you’re figuring it out, then promote it to a plugin the moment you want to reuse it across projects or hand it to someone else.

So when someone asks “should I install this plugin,” the real question underneath is: am I about to hand my agent a bundle of skills, commands, hooks, and outside connections — and do I trust every piece of it? That’s not paranoia. That’s the whole game, and it’s where most plugin advice goes quiet.

The 3 Levels of Trust: Official vs Community vs Random Marketplace

A plugin can contain hooks and MCP servers. Hooks run code on your machine. MCP servers reach out to the internet with your credentials. Installing one is not like adding a browser theme — it’s closer to running a script someone emailed you. Before you type /plugin install, sort the source into one of three tiers.

Tier 1 — Official (Anthropic’s marketplace)

Claude Code adds the official marketplace (claude-plugins-official) automatically the first time you start it. You install from it with a command like /plugin install github@claude-plugins-official, or browse the catalog at claude.com/plugins. This is curated by Anthropic, inclusion is at their discretion, and it’s the only tier I’d install from without reading source first. Even here, I still read what a plugin bundles — but the floor is high.

One correction since I first wrote this: “official marketplace” doesn’t mean “Anthropic wrote it.” When I counted the marketplace manifest on 3 November 2026 it listed 315 plugins, and 40 of them name Anthropic as the author. The rest come from Google, SAP, AWS, Grafana, other vendors and independent developers, and 82 list no author at all. Anthropic decides what goes in, but it doesn’t write most of it. So read the author line in the /plugin panel even inside Tier 1.

Tier 2 — Community (a named person or team you can verify)

A plugin from a marketplace run by a developer with a real GitHub history, a public repo you can read, and a track record is a reasonable bet — after you skim the code. Open the repo. Look at what the hooks actually do and which MCP servers it wires up. Five minutes of reading saves you a very bad afternoon. This is the tier where the honest homework pays off.

Tier 3 — Random marketplace (a URL you found in a thread)

Adding a marketplace is a two-step move on purpose: registering a catalog doesn’t install anything, so you get a beat to think before you pull code down. When the source is an anonymous handle or a marketplace you can’t trace to a human, that beat is your exit. My rule for the fleet is blunt: no unknown-origin plugin touches a container that holds live business credentials. Ever. If I want to try one, it goes on a throwaway machine with nothing to steal. I unpack the full reasoning in my Claude Code security guide for solopreneurs, and it’s the single most-skipped step in every “best plugins” list I’ve read.

The uncomfortable truth: the more useful a plugin is, the more access it wants, and the more access it wants, the more the source matters. Trust is the feature. Convenience is the trap.

5 Claude Code Plugins Worth Installing for a Solo Business

These are the categories I’d actually put on a solo operator’s machine — a one-person show, not a fifty-engineer team. I’m describing capabilities and where they live rather than chasing plugin names that get renamed monthly, because the category is what survives. Every one of these earns its keep by removing a step you currently do by hand.

1. A Git/GitHub workflow plugin. The official GitHub integration is the first thing I install on any machine that touches a repo. It turns “open a PR, write the description, check the CI, respond to review comments” into a conversation. For a solopreneur shipping their own site or product, this is the difference between dreading version control and forgetting it’s there.
2. A code-intelligence (LSP) plugin. The official code-intelligence plugins wire up Language Server Protocol — the same tech behind VS Code’s “jump to definition” and instant error underlines. With it on, the agent sees type errors and broken references the second it edits, instead of confidently shipping something that doesn’t compile. You install the language server binary yourself; the plugin does the wiring. Fewer dumb mistakes, less babysitting.
3. A testing/quality plugin. Anything that gets your agent to run the test suite and read the failures before declaring victory. On the fleet, “green tests or it didn’t happen” is a rule, and a plugin that bakes that loop in means I’m not the one nagging.
4. A documentation or knowledge-base connector. A plugin that gives the agent read access to your own docs or notes so it stops guessing about how your business works. This overlaps with the MCP world — my roundup of the MCP tools that give Claude real superpowers covers the connectors worth the wiring.
5. Your own house plugin. The highest-ROI plugin on my machines isn’t from any marketplace — it’s the one I built for a task I repeat daily. More on that below, because this is where solo operators quietly lap the “download more plugins” crowd.

Notice what’s not on this list: anything that promises to “10x” you with a suite of forty commands you’ll never remember. Five capabilities that each kill a recurring chore beats a mega-pack every day of the week.

The Best Claude Code Plugins, Ranked by What They Actually Run

The section above sorts plugins by job. This one names names. The ranking rule is the one I use for anything that touches a business machine: how much chore does it remove, divided by how much it can do without asking. A plugin made of one slash command that runs only when you call it scores high. A plugin that fires hooks on every edit or sends a token to a remote server has to earn that access. Install counts didn’t decide the order, but they’re in the table so you can see where I disagree with the crowd.

#	Plugin	Publisher	What it installs	Runs code or calls out?	Installs*
1	code-review	Anthropic	1 slash command that runs several review agents with confidence scoring	No hooks, no MCP	438,525
2	commit-commands	Anthropic	3 commands: commit, commit-push-pr, clean_gone	Runs git when you call it	171,244
3	typescript-lsp / pyright-lsp	Anthropic	A language server hookup for your language	Runs the language server binary you install	212,522 / 109,778
4	security-guidance	Anthropic	Warnings when files are edited, an LLM review of the diff at Stop, and a commit reviewer	Yes: 13 hook files on 4 events	241,800
5	claude-md-management	Anthropic	1 command + 1 skill to audit and update CLAUDE.md	No hooks, no MCP	287,247
6	skill-creator	Anthropic	1 skill to write, improve and test your own skills	No hooks, no MCP	385,083
7	github	GitHub (listed by Anthropic)	Remote MCP server at api.githubcopilot.com	Yes: sends your GitHub token	319,381
8	context7	Upstash	Remote MCP server for current library docs	Yes: your queries go to mcp.context7.com	417,801
9	playwright	Microsoft	MCP server that drives a real browser	Yes: opens pages, fills forms	319,887
10	superpowers	Jesse Vincent (obra), partner-reviewed	A large skills library: TDD, debugging, subagent workflows	Pinned to a git commit	1,009,371

*Total installs from a third-party tracker of claude.com/plugins, snapshot dated 18 September 2026. Treat them as rough popularity, not a measure of quality. Publisher and contents were checked against the official marketplace manifest on 3 November 2026.

One-line verdicts

1. code-review — Install first. It catches the problems you’d otherwise ship, and it only runs when you ask.
2. commit-commands — Install. It turns the boring end of every change into one command.
3. typescript-lsp / pyright-lsp — Install the one for your language. Claude sees type errors as soon as it edits.
4. security-guidance — Install if you ship code other people use. Expect it to slow every session a little.
5. claude-md-management — Install. A memory file you don’t prune keeps growing.
6. skill-creator — Install. It’s the first step to building your own plugin later.
7. github — Install if you live in GitHub. Use a token that can only reach the repos you need.
8. context7 — Worth it for fast-moving libraries. Don’t paste private code into its queries.
9. playwright — Situational. Great for testing your own site. Keep it off any machine that’s logged into things.
10. superpowers — Try it, then cut it down. It’s the closest thing on this list to a mega-pack.

The top six are all published by Anthropic, and four of them add no hooks and no MCP server at all. That’s no accident. The best Claude Code plugins for a solo business mostly package instructions. They don’t add new ways to reach the outside world. Install every plugin in the official catalogue with the same command shape:

/plugin install code-review@claude-plugins-official
/plugin install commit-commands@claude-plugins-official
/plugin install claude-md-management@claude-plugins-official

In a session, /plugin install no longer installs straight away. It opens the plugin’s details pane so you can see its skills, agents, hooks and MCP servers and pick a scope (just you, everyone in the repo, or you in this repo only). Read that pane. It’s the five-second version of the vetting I describe above.

Popular, but not on my list

• frontend-design is the most installed plugin on the tracker, with 1,134,112 installs. It’s a single Anthropic skill and it’s good at what it does. But it does UI work, and most of a solo operator’s automation has no UI. Install it the week you build a landing page, not before.
• ralph-loop (196,527 installs) adds a Stop hook that stops the session from ending and feeds Claude’s output back in as the next prompt, up to a max-iterations value you set. That’s clever when you’re sitting at the keyboard. On an unattended machine, a hook whose job is to stop the agent from finishing needs a hard cap you’ve tested.
• code-modernization ships 13 commands, 8 agents and 15 hook files. It may be excellent for a legacy rewrite. It’s also exactly the shape of the “mega-pack” I warn about below.

If an MCP-based plugin is on your shortlist, weigh the token bill too. In MCP vs CLI I measured how much of a raw tool response actually needs to reach the model. Often a CLI the agent already has does the same job for less.

3 Popular Plugins You Can Skip

Now the part the marketplace listings will never tell you: some of the most-installed plugins are dead weight for a one-person business. Not because they’re badly built — because they solve a problem you don’t have, and every plugin you load is more context, more surface area, and more things to reason about. Here’s what I leave off.

1. Giant “everything” mega-packs

The bundle with fifty commands, twelve agents, and a hook for every occasion. It demos beautifully. In daily use it slows your agent down with instructions it’s constantly wading through, and you’ll use maybe four of the fifty. Install the four capabilities you need as separate, small plugins. Lean beats comprehensive when you’re the whole team.

2. Team-collaboration and ceremony plugins

Plugins built around standups, sprint rituals, multi-reviewer approval flows, and handoffs between roles. Genuinely useful at a company. Pure overhead for a solopreneur — you are the standup. Skip anything whose main value is coordinating people you don’t have.

3. “AI-powered” wrappers around something the base agent already does

A plugin that adds a fancy command to “explain this code” or “write a commit message” when Claude Code already does both natively. You’re adding a dependency and a trust decision to buy a feature you already own. Before installing anything, ask: can the base tool already do this if I just ask? Nine times out of ten for these wrappers, yes. If you’re still mapping what the base tool ships with, my breakdown of everything Claude Code actually gives you saves you from re-buying built-ins.

The pattern across all three: bloat rarely announces itself as bloat. It shows up as “more,” and more feels like progress. On a solo machine, subtraction is the upgrade.

Build Your Own Plugin: Turn a Repeated Task Into a Reusable Bundle

This is where the leverage lives, and almost nobody talks about it because it’s less clickable than a “top 10” list. The most valuable thing on my machines isn’t from a marketplace. I built it out of a task I was doing every single day. To be precise: today it’s a skill, not yet a plugin. The difference is the point of this section.

The recipe is embarrassingly simple. You don’t start by writing a plugin. You start by noticing you’ve typed the same request three days running.

1. Spot the repeat. On my fleet it was publishing a blog post: pull the next topic, research the competition, write it, generate images, push to WordPress, set the SEO fields, log it. Same seven-step dance, every day, across every brand.
2. Nail it once in standalone config. Get the workflow working as a plain skill in .claude/ for one project. Don’t package anything yet. Iterate until it’s boringly reliable.
3. Add the manifest. Create .claude-plugin/plugin.json with a name, version, and author. That one file turns your working skill into a real plugin you can drop into any container.
4. Bundle the supporting cast. Fold in the hook that verifies the post went live, the slash command that kicks the whole thing off, and the MCP connection to your CMS. Now the capability travels as one unit.
5. Reuse across the fleet. This is the step I haven’t finished, and that’s the receipt. The publishing workflow that wrote this post is a single skill file (blog-writer.md, 47,730 bytes), and each brand container gets its own copy from a template. It’s one of 39 skill files in this container. Copying files works, but every fix has to land in every copy. When a field-name bug turned up in this one on 6 October, the fix was a hand edit to this container’s copy. Packaging the workflow as one versioned plugin is the fix I’d pick now.

You don’t need to be a developer to do this. If you can describe your process clearly to a smart new hire, you can build a plugin — Claude Code does the packaging while you supply the judgment about what the workflow should be. That’s the whole ballgame for a solo operator: your repeated tasks are the raw material, and a plugin is just the container that makes them reusable.

This is also, honestly, where a done-for-you build earns its money. If you can see the repeated task but don’t have the hours to package it cleanly, that’s precisely the kind of thing I build for operators — book a build session and we’ll turn your weekly grind into a plugin you own.

Why the Container That Published This Post Runs Zero Plugins

Here’s what the machine that wrote and published this article returned at 00:05 on 4 November 2026 (Bali time), Claude Code 2.1.292:

$ claude plugin list
No plugins installed. Use `claude plugin install` to install a plugin.

$ claude plugin marketplace list
Configured marketplaces:
  ❯ anthropic-plugin-directory
    Source: Built in (Anthropic Directory)
  ❯ claude-plugins-official
    Source: GitHub (anthropics/claude-plugins-official)

Two marketplaces are registered and nothing is installed. That’s on purpose, for three measured reasons.

1. Most of the catalogue has never been reviewed

The same container keeps a local cache of the plugin directory. On 3 November 2026 it held 3,604 listings. 3,162 of them (87.7%) show a review state of “none”. 350 were reviewed by a partner and 92 by Anthropic. That isn’t a criticism of the directory, which is open on purpose. But “it’s in the directory” tells you almost nothing about safety. On a laptop that’s a judgment call. On a server that holds WordPress, Airtable, Gmail and payment-adjacent credentials for ten businesses, it’s a hard no unless someone has read the code.

2. An unattended agent can’t read the details pane

The install flow assumes a human looks at the components before choosing a scope. My containers run on a cron schedule with nobody watching. Anything that adds hooks or MCP servers there is a change to production. It gets reviewed and pinned like any other code, not installed because a session suggested it.

3. Every plugin is context the agent pays for on every run

Each run already loads a 47,730-byte skill file and a 29,597-byte CLAUDE.md before it does any work. I broke that file down line by line in CLAUDE.md best practices. A plugin’s skills, agents and tool definitions add to that bill, and the agent has to consider them on every turn. The docs now give you a “Not used recently” header in the /plugin Installed tab, which is the vendor quietly admitting the same thing.

So the rule for the fleet is simple. Interactive machine: the top six from the ranked list, all at user scope. Production container: nothing from a marketplace. Anything production needs is a skill or hook I’ve read line by line, and when it has to travel between machines it should become a private plugin pinned to a commit. Same capability, different trust boundary.

Plugins vs MCP Servers vs Skills: When to Reach for Each

These three get blurred together constantly, and the confusion costs people time. Here’s the clean mental model, operator-style.

• A skill is knowledge and instructions — the “how we do this task” write-up the agent loads when it’s relevant. Reach for a skill when the thing you’re adding is judgment or a procedure: our brand voice, our invoicing rules, the steps to onboard a client.
• An MCP server is a connection to an outside system — a door to your database, your Notion, your Stripe, your calendar. Reach for MCP when the thing you need is access to data or an action in another tool. If you’re fuzzy on the concept, my plain-English explainer on what an MCP server actually is clears it up in five minutes.
• A plugin is the shipping container — it bundles skills, commands, hooks, and MCP connections so a whole capability installs as one unit. Reach for a plugin when you want to package and reuse a combination of the above, especially across projects or with other people.

The one-liner I use: skills are what your agent knows, MCP is what your agent can reach, and plugins are how you box that up and ship it. You don’t choose between them — a good plugin usually contains a skill and an MCP server. You choose based on whether you’re teaching, connecting, or packaging. Get that straight and 90% of the “which one do I use” paralysis disappears.

FAQ: Cost, Vetting, and Uninstalling Safely

The questions I get asked most about Claude Code plugins, answered straight.

What are the best Claude Code plugins to start with?

For most solo operators: code-review, commit-commands, the LSP plugin for your main language, and claude-md-management. All four are published by Anthropic, and three of them add no hooks and no MCP servers. Add security-guidance once you ship code other people rely on. Add an MCP plugin like github or context7 only when you need the outside system it connects to. The ranked table above has the full ten.

Do plugins cost extra money?

The plugin system itself is free — installing from the official marketplace doesn’t add a fee. What can cost you is what a plugin does: a plugin that makes lots of model calls uses tokens against your plan, and one that connects to a paid third-party service means that service’s bill, not Anthropic’s. Your Claude Code subscription is the same either way. If you’re weighing that up, I broke down the real numbers in my honest breakdown of whether Claude Code is free.

How do I vet a plugin before installing it?

Three checks, in order. One: where’s the source — official, a verifiable person, or an anonymous URL? Two: open the repo and read what the hooks run and which MCP servers it connects to; you’re looking for anything that reaches out to the network or touches credentials. Three: does the base tool already do this? If the answer to that last one is yes, don’t install. When in doubt, try it on a machine that holds nothing valuable first.

How do I uninstall a plugin safely?

Use /plugin uninstall name@marketplace in a session, or claude plugin uninstall name@marketplace from your shell (add --scope project if the repo enabled it). If a project enables the plugin, Claude Code asks whether to disable it just for you or remove it for everyone. Two cleanup steps people miss. First, dependencies a plugin pulled in stay installed until you run claude plugin prune. Second, if the plugin added an MCP connection tied to an API key or token, revoke or rotate that credential too. Uninstalling removes the wiring, but it doesn’t take back access you already granted. Treat a plugin’s credentials the way you’d treat a departing contractor’s and cut them off properly.

Can plugins slow my agent down?

Yes, and it’s the most underrated cost. Every plugin adds instructions and tools the agent has to consider on every turn. A handful of focused plugins is invisible; a stack of mega-packs makes the agent slower and more distractible. Prune ruthlessly. If you haven’t used a plugin in a month, remove it. The /plugin Installed tab now groups those under a “Not used recently” header, so this takes about thirty seconds.

Plugins or standalone config — which should I start with?

Start standalone in .claude/ while you’re experimenting; it’s faster to iterate. Promote to a plugin the moment you want to reuse the capability across projects or share it. That’s not just my rule — it’s the official guidance, and it’s the right one.

The Operator’s Bottom Line

Claude Code plugins aren’t a magic app store, and the goal was never to install the most of them. A plugin is a bundle of skills, commands, hooks, and connections — powerful exactly because it can run code and reach your data, which is the same reason you vet the source before you trust it. Install the few that kill a recurring chore, skip the mega-packs and the ceremony, revoke credentials when you uninstall, and — this is the part that actually moves the needle — build your own out of the task you already do every day. That last move is how a one-person business starts running like a ten-person one. It’s the whole reason I can operate a fleet of them solo, and it’s more within reach than the “top 10” lists want you to believe.