Claude Connectors & Plugins — the full directory

Three prompts, and Claude does the rest

I call them connectors in the reel because that is what they feel like. Their real name is MCP servers — little plugs that give Claude a new ability, like driving a browser or reading your Notion. You do not need to understand the plumbing. You need the right five and a prompt that installs them.

Open Claude Code in your terminal and paste whichever of these three fits. It installs one at a time, checks each one actually connected, and stops to ask you when it needs a key.

1 · Install the classic five

Install five MCP servers into my Claude Code setup. Do them one at a time, and after each one run `claude mcp list` and confirm it says Connected before you move to the next.

If a server needs an API key, STOP and ask me for it. Tell me the exact URL to get it and the menu path once I am there. Wait for me to paste it, then carry on.

The five:

1. Playwright - browser control, built by Microsoft, no key needed:
claude mcp add playwright -- npx @playwright/mcp@latest

2. Sequential Thinking - structured reasoning, Anthropic's own, no key needed:
claude mcp add sequential-thinking -- npx -y @modelcontextprotocol/server-sequential-thinking

3. Context7 - current documentation for any library. Ask me for a free key from context7.com/dashboard first, then:
claude mcp add --transport http context7 https://mcp.context7.com/mcp --header "Authorization: Bearer MY_KEY"

4. Firecrawl - scrape and crawl any website. Ask me for a free key from firecrawl.dev first, then:
claude mcp add firecrawl --env FIRECRAWL_API_KEY=MY_KEY -- npx -y firecrawl-mcp

5. Perplexity - live web research with citations. Ask me for a key from console.perplexity.ai first, then:
claude mcp add perplexity --env PERPLEXITY_API_KEY="MY_KEY" -- npx -y @perplexity-ai/mcp-server

One thing to check before you start: if `claude mcp list` already shows a server called `server-perplexity-ask`, remove it. That package has not been updated since April 2025 and the line above replaces it.

When all five are done, run `claude mcp list`, show me the output, and give me one real test prompt for each. If anything failed to connect, diagnose and fix it before you finish.

2 · Work out MY five, then install them

I want to add MCP servers to Claude Code, but only the ones I will genuinely use. Every connected server sits in my context on every message, so more is not better.

First, ask me these four questions one at a time and wait for each answer:
1. What do I spend most of my working week actually doing?
2. Which apps do I have open right now?
3. Do I build or edit websites, or not really?
4. Am I happy signing up for free API keys, or do I want no-signup ones only?

Then run `claude mcp list` so you know what I already have.

Then recommend exactly five servers. For each one, tell me in a sentence why it earned its place. Then tell me which ones you deliberately left out and what I am giving up by not having them. If two of your five do overlapping jobs, drop one and say which.

Wait for me to approve the five before you install anything.

Then install them one at a time, checking each with `claude mcp list` before the next, pausing to ask me for any API key with the exact URL to get it.

Finish with one test prompt per server, written for my actual work rather than a demo.

3 · Audit what I have already got

Audit my MCP setup and help me cut it back.

Run `claude mcp list` and show me everything connected. For each server, tell me in one line what it is for and roughly how many tools it exposes.

Then ask me which ones I have actually used in the last month. For anything I have not, give me the exact `claude mcp remove` command.

Flag anything that is:
- doing the same job as another one I have (two search servers, two browser servers)
- installed but failing to connect
- from a package that has not been updated in over a year

Do not remove anything yourself. Give me the commands and let me decide.

If Claude asks permission, that is normal

Claude Code checks with you before running commands on your machine. When it asks "can I run this?", that is it being polite, not a sign something has gone wrong. Say yes and let it keep going.

The five plugins — and where the rest live

If you landed here from the plugins video, this section is yours. I promised you my setup for the five, plus twenty more. Here is the setup, and the rest are split between this section and the connector directory further down.

Before the list, two things I got wrong on camera. I would rather correct them here than let you find out at the terminal.

Correction 1: three of the five are not actually plugins

I called them all plugins. Only frontend-design and Playwright install with /plugin install. OmniRoute is a separate program you install with npm. Headroom and MemPalace both technically offer a plugin, but the plugin on its own does nothing — you have to install the actual tool first, and then the plugin just wires it in.

This matters because if you type /plugin install omniroute you will get an error and assume you have done something wrong. You have not. Each card below has the command that genuinely works.

Correction 2: the "1.6 billion free tokens" number

I said OmniRoute includes over 1.6 billion free tokens a month. When I filmed, their README did say roughly 1.6 billion. On 08/09/2026 the version you would install today says ~1.51 billion, and the next one already says ~1.47 billion. It moves, and it moves down.

The bigger issue is the word included. Nothing is included. It is a running tally of what you would get if you personally signed up to about 34 different providers' free plans and let OmniRoute route between them. Roughly two thirds of that headline figure is a single provider's free tier. Their own documentation says the numbers "are re-audited every two weeks" and "move both ways".

It is still a real thing and it can genuinely keep you working when you hit a wall. Just do not sign up expecting a billion tokens to be sitting there waiting for you.

Plugin, connector, or program?

Three different things get called "plugins" and they install in three different ways. Two minutes here saves you an afternoon.

🧩

A plugin changes Claude

Instructions, skills and hooks that change how Claude behaves — how it designs, what it checks, when it stops. Installs with /plugin install name@marketplace, inside Claude Code.

🔌

A connector gives Claude access

Also called an MCP server. It hands Claude something it did not have — the live web, a browser, your calendar. Installs with claude mcp add .... All 25 in the directory below are these.

🔧

A program sits underneath

Software that runs on your computer alongside Claude Code. Installs with npm, uv or pip in your terminal, like any other app. OmniRoute and Headroom are these.

All of this runs on Mac, Windows and Linux

Every single item on this page. Claude Code itself supports macOS, Windows and Linux, and I checked each tool's own installer list rather than assuming. Where something has a genuine platform quirk, it is written on that item's card.

The five from the video

Each one checked at its own repository on 08/09/2026 — the install command is copied off the project's own instructions, not written from memory.

OmniRoute

In the videoNot a plugin

Keeps you working when you hit a usage limit. It sits between Claude Code and a pile of model providers, and when one runs out it routes you to another one mid-task instead of stopping you dead. Their own README describes it as "quota-aware auto-fallback" across four tiers — your subscription first, then paid API, then cheap models, then free ones. That part is real and it is the reason it is on the list.

npm install -g omniroute

What it will not do: install as a plugin — there is no plugin, and /plugin install omniroute does not exist. It is a program you install with npm, run with omniroute, and then point Claude Code at using omniroute configure claude. It needs Node 22.22.2 or newer, and it runs on Mac, Windows and Linux. And about the free tokens: I said "over 1.6 billion, included" on camera. Both halves have a problem — see the box directly below. One more thing worth knowing: their own free-tier file marks thirteen of those providers as ToS-risky — one of them says outright that automated querying is against its terms — and flags that the marking is "advisory, not a routing gate", so those providers are still used by default. If you are doing client work, go into the dashboard and turn the flagged ones off rather than leaving it on the defaults.

frontend-design

In the videoNo keyAnthropic

The one on this list that is genuinely, straightforwardly a Claude Code plugin, published by Anthropic and marked Anthropic Verified. It is a single instruction file that changes how Claude designs. It tells Claude to commit to an actual aesthetic direction instead of hedging, and it names the tells it wants avoided — "generic fonts like Arial and Inter", and specifically purple gradients on a white background.

/plugin install frontend-design@claude-plugins-official

What it will not do: hire you a designer, and I want to be straight about that because I oversold it in the video. There is no component library in here, no design system, no assets — it is a prompt. A good one, and the difference is obvious on the first page you build. But what it really does is stop Claude reaching for the same default look every time. Nothing to sign up for, works on any operating system, and Anthropic's marketplace registers itself the first time you open Claude Code, so there is no marketplace to add.

Headroom

In the videoNeeds the CLI first

Squeezes the stuff Claude reads before it gets to the model. Their words: "Headroom compresses everything your AI agent reads — tool outputs, logs, RAG chunks, files, and conversation history — before it reaches the LLM." All of the compressing happens on your own machine; nothing gets sent anywhere to be shrunk.

uv tool install --python 3.13 "headroom-ai[all]" && headroom wrap claude

Two corrections to what I said. I described it as reading everything at the start of a session and clearing out the junk. It is not a session-start clean-up — it works continuously, on every payload, all session. And "the same output for much cheaper" is too strong: their own benchmark table shows 21% saved on code search and 42% on codebase exploration, and the repo's own summary is "20% fewer tokens for coding agents, 60-95% fewer tokens for JSON". So for the work you will actually be doing, expect roughly 20-40%, not 90%. Other things worth knowing: it publishes installers for Windows, Linux and both Apple Silicon and Intel Macs. headroom wrap also installs a second tool called Serena across all your projects unless you add --code-memory none. And it downloads a model the first time it runs, which will fail on a locked-down work network. The command above needs uv installed first — ask Claude to install it for you, it takes a few seconds and it knows which command your computer needs.

MemPalace

In the videoBeta

Gives Claude a memory that survives between chats, so you stop re-explaining your business every time. It stores what you tell it word-for-word and searches it by meaning — people and projects become wings, topics become rooms. It is local-first by default: "Nothing leaves your machine unless you opt in." No API key needed for the standard setup.

npx skills add MemPalace/mempalace

Read this before you install it. Their own README carries a warning in capitals about fake versions: "Beware of impostor sites. MemPalace has no other official websites... Any other domain is an impostor and may distribute malware." Only ever go to github.com/MemPalace/mempalace. There is also a trap in the setup: "Claude Code sessions expire in 30 days without auto-save hooks wired." If you install it and skip the hooks, you will think it is remembering and it will not be. The command above is the guided setup, which wires them for you. It is also officially still in beta and only five months old — fine to try, not something to build a business process on yet.

Playwright

In both videosNo keyMicrosoft

Gives Claude its own browser. It clicks, fills in forms, takes screenshots, reads the console, and checks whether the thing it just built actually works. This is Microsoft's official one, and it is already number 6 in the connector directory below — it is the one item that appears in both videos.

claude mcp add playwright npx @playwright/mcp@latest

What it will not do: stay out of sight. It runs headed by default, which means a real browser window opens on your screen and moves by itself. That is it working, not something going wrong, but it startles people the first time. It needs Node 18 or newer, and runs on Mac, Windows and Linux. It is also still on version 0.0.80 — heavily used, but not yet a stable 1.0, so expect the odd change. "It fixes its own work" is fair, with one caveat: the fixing is Claude reading the browser's errors and having another go. The connector supplies the eyes, not the judgement.

Or let Claude do it

This installs the two that are genuinely plugins and the browser connector, checks your setup, and stops. It deliberately does not touch OmniRoute, Headroom or MemPalace — those three change how your whole setup runs, and you should install them one at a time, deliberately, having read their cards.

4 · Set up the safe three for me

I want to install some Claude Code plugins. Work out which operating system I am on first and use the right commands for it — do not assume.

Here is what I want, in this order. Stop after each one and tell me plainly whether it worked before you move on.

1. Anthropic's official plugin marketplace should already be registered. Check with /plugin marketplace list. If it is not there, add it with: claude plugin marketplace add anthropics/claude-plugins-official

2. Install the frontend-design plugin from it: /plugin install frontend-design@claude-plugins-official
   Then tell me in one sentence what it will change about how you build pages.

3. Install the Playwright browser server: claude mcp add playwright npx @playwright/mcp@latest
   Warn me first that by default this opens a real, visible browser window on my screen when you use it. That is normal, not a problem, but I want to know before it happens.

4. Tell me whether I have Node installed and what version. Playwright needs Node 18 or newer.

Then STOP. Do not install anything I have not asked for, do not add any other marketplace, and do not edit my settings files beyond what these commands do themselves.

Finally, write me a short note in this folder called claude-plugins.md listing what you installed, the exact command for each, and how I would remove each one later.

Eight more, already on your machine

Anthropic's own marketplace registers itself the first time you open Claude Code, so these need no setup at all — one command each. These are the eight I would actually pick out of it for someone running a business rather than a codebase.

skill-creator

AnthropicAlready there

Build your own skills — it interviews you, writes the skill and tests it. The one to install if you keep pasting the same long instructions.

/plugin install skill-creator@claude-plugins-official