Claude Code plugins: the 10 GitHub repos that actually change something

A Claude Code plugin is an installable package that adds skills, commands, agents and MCP servers to your session in one go. One command, and Claude can do one more thing. Anthropic's official marketplace listed more than 300 of them in October 2026, on top of the repositories published straight to GitHub by vendors and developers.

Here are the ten I install and recommend in training, with what each one brings, its install command and its limits. Moving to Claude Code is already a step. Equipping it is the next one, and that is what separates a demo tool from a production tool.

Plugin, skill or MCP server, what you are actually installing

Three words circulate in the artificial intelligence ecosystem and get mixed up. A skill is a folder of instructions, scripts and resources that Claude loads when the task calls for it. An MCP server is access to an outside tool, letting Claude read and write in your CRM, your accounting system or your database. A plugin is the package that carries both, plus commands and agents.

A skill loads in three stages, described by Anthropic in its October 2025 engineering post on Agent Skills. At startup only the metadata enters the context, roughly a hundred words per skill. The body of the instruction file arrives only when a task triggers it. Linked resources are read only when genuinely needed. The format became an open standard in December 2025.

This is exactly the mechanism I described in my newsletter on Claude Code. Without skills, you pasted the same meta-prompt into every conversation, the context blew up and you paid to re-explain the same thing over and over. At Tandem, each skill replaces a playbook we used to paste by hand, from commercial proposals to end-of-engagement reports.

How I picked these ten repos

The criteria come before the names, so you can argue with the ranking.

• Useful outside development. A repo that only helps write code excludes nine business users out of ten.
• Actively maintained. All ten received commits in the month before this October 2026 check.
• One-command install. A repo that demands thirty minutes of setup will never be adopted by a team.
• Identifiable publisher. Anthropic, Microsoft, HeyGen, Apify, or a developer whose name is on the line.

Full disclosure on what I have actually used. Anthropic's document skills, HyperFrames, Firecrawl and Playwright run in our projects. Superpowers, Humanizer, OpenSEO and Matt Pocock's skills, I tested on internal work. Apify and Addy Osmani's skills, I rely on their documentation and their code, with no client deployment behind them.

1. Anthropic skills, office documents that finally look right

The anthropics/skills repository gathers the reference skills published by Anthropic, including the ones that produce Claude's Word, PowerPoint, Excel and PDF files. These are the same implementations used in production inside the product. It passed 180,000 stars in October 2026, which makes it the obvious entry point.

• Office files generated with real layout, where an exported Markdown stays raw.
• The Agent Skills format specification, useful if you write your own.
• A skill template ready to duplicate for your own procedures.

Install. Add the marketplace with /plugin marketplace add anthropics/skills, then install document-skills@anthropic-agent-skills. The demo skills come through example-skills.

Limits. Anthropic states that these skills serve demonstration and education, and that the behaviour you get may differ from the product. The document skills are source-available, under a more restrictive licence than the rest.

My take. This is the first one I install on any new machine, and the one with the most immediate effect on non-technical teams. A meeting report that comes out as a clean deck convinces faster than a code generation demo. Keep in mind that your brand styling is still on you, through a rules file in the project.

2. Superpowers, method before keyboard

Superpowers is a full development methodology delivered as composable skills. The agent starts by extracting a spec from you, cuts it into readable chunks, then produces an implementation plan before working task by task, tests first. The repository passed 300,000 stars in October 2026.

• Enforced framing, with questions asked before a single line is written.
• Subagent-driven development, with code review built in.
• Systematic debugging rather than blind fixes.
• A skill that teaches Claude to write and test new skills.

Install. A single command, /plugin install superpowers@claude-plugins-official, since the plugin sits in Anthropic's official marketplace.

Limits. The method is demanding and stretches the start of every piece of work. On a three-line fix it will cost you time. It targets projects that run for days.

My take. I tested it on internal work and the gain shows mostly in what the agent stops doing. The endless back and forth disappears because the need was settled up front. If you install one method plugin, take this one.

3. Apify MCP, thousands of scrapers without writing a line

The Apify MCP server gives Claude access to Apify's catalogue, meaning thousands of ready-made scrapers and automations for social networks, search engines, maps and online retail. You describe the data you want, Claude picks the actor and brings back the result.

• Extraction from Google Maps, Instagram or a marketplace, with no code.
• A connection by URL alone, with delegated authentication.
• A hosted server, which avoids installing anything locally.

Pricing. The free tier opens roughly 5 euros of usage a month, with no card required. The first paid tier sits around 18 euros a month, checked with the vendor in October 2026, and billing follows actual usage.

Limits. Cost climbs fast at volume, because you pay per run. Quality depends on the actor you pick from the catalogue, and they are not all equal.

My take. I have not deployed it on a client engagement, so I stick to what the documentation and the code allow me to verify. The pricing model deserves attention. An agent firing scrapers in a loop produces an invoice before it produces a result, so set a cap before you loosen the reins.

4. HyperFrames, video written in HTML

HyperFrames, published by HeyGen, turns HTML, CSS and animation into deterministic MP4 files. Claude writes the composition like a web page, the engine renders it as video. The project passed 74,000 stars in October 2026 and sits in the official marketplace.

• Product videos, animated explainers and title cards generated from a prompt.
• Captions, voiceovers and audio-reactive visuals.
• Deterministic rendering, so results repeat from one run to the next.

Install. The plugin installs through /plugin install hyperframes@claude-plugins-official. Local rendering needs a recent Node version.

Limits. You stay in motion design and graphic treatment. For shooting and editing real footage, look elsewhere.

My take. We use it for the short formats that go with our content, and the gain lies in repeatability. Once the brand rules are written into the composition, the tenth video comes out like the first, with no colour or typography drift.

5. Matt Pocock's skills, the need questioned first

Matt Pocock's skills aim at real engineering rather than improvised code. Their most telling command grills you on the need before anything gets written, forcing you to spell out what you actually want. The repository passed 300,000 stars in October 2026.

• Short, editable skills, designed to be adapted to your context.
• A spec and ticket flow, then test-driven development.
• Code review and domain modelling.

Install. Through claude plugin install mattpocock-skills@claude-plugins-official, from your shell or from a session.

Limits. The approach partly overlaps Superpowers and Addy Osmani's skills. Installing all three stacks up three competing methods, and Claude will hesitate.

My take. I tested it internally and the difference with Superpowers comes down to size. Here the skills are short enough to read in full and rewrite to your taste, which matters when a team wants to understand what it runs.

6. Humanizer, AI writing tells stripped from a text

Humanizer rewrites generated text so it reads like something a person wrote, without changing what it says. The skill builds on the guide to signs of AI writing maintained by Wikipedia editors, and hunts twenty-six patterns, among them the negation followed by a correction, the one-line closer and the hollow maxim.

• A pass over an existing text, without rewriting the substance.
• An explicit, auditable list of patterns rather than a black box.
• Identical behaviour on Claude Code and on other agents that accept skills.

Install. Two commands, /plugin marketplace add blader/humanizer then /plugin install humanizer@humanizer.

Limits. The author says it himself, he wrote the skill for human readers and most detectors still flag the output. The gain is about reading comfort.

My take. I tested it on internal drafts and it catches turns of phrase I let through. It works on the form only. A text with no angle comes back clean and still has no angle, so the substance work happens first.

7. Firecrawl, the web turned into usable data

Firecrawl turns any website into clean Markdown or structured data, page by page or site-wide. The project passed 196,000 stars in October 2026 and its plugin sits in the official marketplace. It also ships an agent that can search, navigate and extract across several sources from a plain description.

• A page converted into clean text, stripped of navigation and banners.
• A whole site crawled, with structured extraction against a schema you supply.
• Built-in web search, useful for competitive research or sourcing.

Pricing. The free tier opens about a thousand pages a month. The first paid tier sits around 18 euros a month, checked with the vendor in October 2026, and the team tier around 90 euros.

Limits. Sites behind bot detection hold out, and credit burns fast on a large domain. Check the terms of use of the sites you crawl.

My take. This is the one that most changes the nature of the work you can hand to Claude. As long as an agent only reads what you paste, it stays a desk assistant. Plugged into the web it becomes a research tool, and our sector monitoring runs on it.

8. Addy Osmani's Agent Skills, the full cycle in nine commands

Addy Osmani's Agent Skills cut the development cycle into nine commands, from spec to ship, through planning, slice-by-slice building, testing, review and web performance audit. The repository passed 111,000 stars in October 2026.

• One command per stage, which makes the process legible to a non-developer.
• An automatic mode that chains every task after a single plan approval.
• A built-in web performance audit, rare in this kind of collection.

Install. Through /plugin marketplace add addyosmani/agent-skills, then installing the plugin. Commands that share a name with a built-in command are called in their namespaced form.

Limits. The repository is not in the official marketplace, so there is no auto-update and no context cost estimate. And it largely overlaps the two methods above.

My take. Here I rely on reading the code and the documentation, with no client deployment. The value lies in the granularity of the commands. A team that rejects a full method can take only two of them, framing and review, and keep its habits for the rest.

9. OpenSEO, search tracking without the monthly subscription

OpenSEO is an open alternative to the classic search suites, with keyword research, rank tracking, competitor analysis, backlinks and site audits. It exposes an MCP server, so Claude Code queries your search data inside the session.

• Keyword research and rank tracking driven from the agent.
• Hosting on your side, by container or on your own Cloudflare account.
• Usage-based billing, since you bring your own data provider key.

Pricing. The vendor-hosted version sits around 9 euros a month, checked in October 2026. Self-hosted, you only pay for the data you consume at the provider.

Limits. Without a data provider key, the tool returns nothing. History depth stays far behind the established suites, and setup takes a real hour.

My take. I tested it on internal work, and the real value lies in agent access rather than in the interface. Asking Claude which queries leave a page stuck on page two, then having it fix the page straight away, removes the export and the copy-paste between two tools.

10. Playwright CLI, your pages tested in a real browser

Microsoft publishes playwright-cli, a command-line interface shipped with its own skills. Claude drives a real browser, fills a form, clicks, takes screenshots and checks a journey. The vendor's argument is about context cost, because a short command consumes far less than a full accessibility tree.

• A user journey replayed end to end, with screenshots as proof.
• Low context consumption, which leaves room for the code.
• Installation through the package manager, independent of marketplaces.

Install. Install the package globally, then run playwright-cli install --skills. The official marketplace offers the Playwright plugin in its MCP form, which answers the same need at a higher context cost.

Limits. The subject stays technical, and it targets teams that ship interfaces. An office user will get nothing out of it.

My take. We use it to check our own pages, and it settles the killer question in acceptance testing. Claude claims the form works, the screenshot proves the submission goes through. Between the two versions, take the command line if your context is already loaded.

How to install a Claude Code plugin in one command

Claude Code registers Anthropic's official marketplace by itself the first time you start an interactive session. For a plugin that sits there, one command is enough, for example /plugin install superpowers@claude-plugins-official. The plugin's details pane opens first, listing what will be installed, and the install waits for your choice.

For a repository published elsewhere, add its marketplace first with /plugin marketplace add followed by the owner and the repository, then install with the plugin name followed by the marketplace name. From a script, the same actions exist as terminal commands prefixed with claude plugin.

The risk nobody checks before typing the command

Anthropic puts it bluntly in its documentation, an installed plugin can run arbitrary code on your machine with your user privileges. A plugin's hooks, monitors and MCP servers run outside the sandbox, and Claude Code's permission rules stop at the tool calls Claude makes.

The risk is documented. In September 2025, the security vendor Snyk analysed a malicious MCP server published on the npm package manager. The package sent email on behalf of an assistant, and added a blind copy to an outside address on every outgoing message, attachments included. The behaviour arrived through an update, after several clean versions.

The check comes down to four points, and it takes two minutes.

1. Who publishes it. A marketplace name tells you who runs the catalogue. It says nothing about what each plugin does. The official names are reserved for Anthropic's repositories, so nobody can pose as them.
2. What will be installed. The details pane lists commands, agents, skills, hooks and servers. It flags that a hook exists, without saying what it runs.
3. What the hook runs. Open the repository and read the hooks file, the MCP configuration file and the contents of the executables folder. This is the step everyone skips.
4. What changes afterwards. Auto-update rewrites on disk the files you reviewed. It is on by default on the official marketplace, and off elsewhere.

For a team, the question is settled by an approved list rather than by trust. Managed settings allow or block marketplace sources and force the plugins you picked. We cover this in AI agent security, because a plugin dropped on a laptop grants the same access as an agent deployed on a server.

Which repo to start with, by role

• You do not code. Anthropic's document skills, then Humanizer if you write a lot. Those two produce a visible result within the first hour.
• You develop. Superpowers for method, Playwright for verification. One method at a time, or Claude arbitrates between competing instructions.
• You work with data or research. Firecrawl first, Apify next if the scraper catalogue is useful to you.
• You work in marketing. OpenSEO for search, HyperFrames for animated formats.

Which Claude Code plugins to install first?

Start with three. Ten at once and nobody opens them. Anthropic's document skills for office work, Superpowers for method, Firecrawl for web access. Those three cover most of a team's needs, and all of them sit in a catalogue run by an identified publisher.

The rest comes from real need, never from curiosity. An unused plugin costs context on every message, so it costs room and money. Claude Code can list the plugins you have not used recently, and that is the first thing to check when a session turns slow.

The model is the same for everyone. Equipment makes the difference, and it gets decided as a team rather than laptop by laptop. At Tandem it has become the central topic of our Claude Code training, across more than 50 companies. If you are new to the tool, the guide for non-developers comes before this one.