Claude Code out of the box can only talk. These five MCP servers give it live web access, whole-site crawling, image and video generation, and full browser control. Each one installs with a single command in under ten minutes.
Claude Code writes you a plan. It answers your question. Then it stops.
It cannot open the page you asked about. It cannot test the app it just built. It cannot pull the images your post needs.
Out of the box, Claude is a very sharp chat box with no hands. MCP servers are the hands. Each one plugs a real capability into Claude, and most take one line to add. Here are the five that matter, what each is for in a real business, and the exact command to install it.
What Is an MCP Server?
An MCP server is a small connector that gives an AI agent a new ability: reading a live web page, driving a browser, calling an image model. MCP stands for Model Context Protocol, an open standard, so the same server works in Claude Code and in Codex, not one tool only. You install it once. From then on the agent can use that ability in any session, without you wiring it up again.
That is the whole shift. A skill teaches Claude a procedure it already had the parts for. An MCP gives it a part it never had. If you want the procedure side, the five skills that ship real work is the companion to this one. That guide is what Claude knows. This one is what Claude can touch.
How To Read The Install Commands
Every command below is one line you paste into your terminal. claude mcp add registers the server. The npx part downloads and runs it. Two of the five want an API key from the tool's own site, and the command shows exactly where it goes.
You run each line once. Claude remembers the connection after that. To confirm a server is live, ask Claude to list its MCP servers, or type claude mcp list.
1. Perplexity MCP: Live Web Access Inside Claude
Claude's training has a cut-off date. Perplexity does not. This server gives Claude live internet access directly inside your workflow, so it can check a fact, read a current price, or pull today's news before it writes a word.
For an operator this is the one you feel every day. Confirm a supplier claim before it goes in a client email. Check a competitor's live pricing page. Pull the latest number instead of the number Claude half-remembers from last year.
Copy the install command:
Copy this.
claude mcp add perplexity --env PERPLEXITY_API_KEY="your_key" -- npx -y @perplexity-ai/mcp-server
Repo: github.com/perplexityai/modelcontextprotocol
Your first real run: ask Claude something it cannot know from memory. "Check the current pricing on [A COMPETITOR]'s website and tell me their cheapest plan." Watch it go out to the live web and come back with a real answer.
The mistake that makes it fail: skipping the key. Perplexity needs an API key from your Perplexity account pasted where the command says your_key. Leave it blank and the server loads but every request comes back empty. Get the key first, then run the line.
2. Firecrawl MCP: Crawl a Whole Site Into the Conversation
Perplexity fetches an answer. Firecrawl fetches everything. It crawls entire websites and loads the content straight into your conversation for research or RAG workflows.
Point it at your own documentation, a competitor's full site, or a supplier's catalogue, and Claude reads all of it, not one page. That is how you turn a sprawling website into something Claude can actually reason over: summarise a rival's whole offer, pull every price across forty pages, build a knowledge base from your own help centre.
Copy the install command:
Copy this.
claude mcp add firecrawl --env FIRECRAWL_API_KEY="your_key" -- npx -y firecrawl-mcp
Repo: github.com/firecrawl/firecrawl-mcp-server
Your first real run: "Crawl [YOUR OWN WEBSITE] and list every service we mention, with the page each one lives on." You get a map of your own site in one pass, and you usually find something you forgot was there.
The mistake that makes it fail: pointing it at a giant site with no limit. Firecrawl is powerful and it bills per page crawled. Start with one section or set a page cap. Do not aim it at a thousand-page store on the first try and wonder where your credits went.
3. Glif MCP: Hundreds of Image and Video Models on Tap
Glif connects Claude to hundreds of image and video generation models and optimizes your prompts automatically. You describe what you want in plain words. Glif picks a model, sharpens the prompt, and returns the asset.
For a business that ships content, this closes the gap between "write the post" and "make the visual". Generate a header image for an article. Draft a product shot. Rough out a short video clip. All without leaving the session where you were already working.
Copy the install command:
Copy this.
claude mcp add glif -- npx -y @glifxyz/glif-mcp-server
Repo: github.com/glifxyz/glif-mcp-server
Your first real run: "Make me a clean, on-brand header image for a post about hiring your first virtual assistant." See what comes back, then refine it in the same thread.
The mistake that makes it fail: treating the first image as final. The value is the loop. Tell Claude what is wrong in plain words ("too busy, drop the background, make it lighter") and let it re-run. One prompt rarely lands. Three usually do.
4. Playwright MCP: Let Claude Drive a Browser
This is the one that turns Claude from a writer into a worker. Playwright lets Claude control a browser: test apps, upload files, fill forms, and interact with websites for you.
The ops uses are direct. Test your own booking flow end to end before a campaign goes live. Fill a repetitive web form. Check that a client's site still works after a change. Anything a person would do by clicking, Claude can do here, and it reports what it saw.
Copy the install command:
Copy this.
claude mcp add playwright -- npx -y @playwright/mcp@latest
Repo: github.com/microsoft/playwright-mcp
Your first real run: "Open [YOUR BOOKING PAGE], go through the whole booking as if you were a customer, and tell me every step and anything that breaks." You get a real walk-through of your own funnel from the outside.
The mistake that makes it fail: vague instructions. Playwright does exactly what you say, so "test the site" gets you a shrug. Name the page, the path, and what a pass looks like. If the run goes wrong, tell it the literal fix ("you clicked the wrong button, use the one labelled Book Now") and it corrects. Playwright drives a fresh, clean browser, which is why the next one exists.
5. Chrome MCP: Work With What You Already Have Open
Playwright opens a blank browser. Chrome MCP works with the one you are already in. Inside Claude Code, it lets Claude interact with whatever you currently have open in Chrome, logged-in tabs and all.
That is the difference between a demo and your actual work. Your real inboxes, your real dashboards, the tools you are already signed into. Claude can act inside them where a fresh browser would hit a login wall.
Copy the install steps:
Copy this.
npm install -g mcp-chrome-bridge
Then download the extension from the repo's Releases, open chrome://extensions/, turn on Developer mode, and load the unpacked extension.
Repo: github.com/hangwin/mcp-chrome
Your first real run: with a tab already open on a tool you use, ask Claude to read what is on screen and pull out the piece you need. It works against your live session, not a copy.
The mistake that makes it fail: forgetting it acts as you. Because it uses your open, logged-in Chrome, it can do anything you can do in those tabs. Treat it like handing someone your unlocked laptop. Watch what it does the first few runs, and do not point it at anything you would not click yourself.
Which MCP For Which Job
When two of these blur together, this is the split.
| You want to... | Install |
|---|---|
| Check a live fact, price, or today's news | Perplexity |
| Load an entire website into the conversation | Firecrawl |
| Generate images or video from a description | Glif |
| Test an app or fill a form in a clean browser | Playwright |
| Act inside tabs you are already logged into | Chrome MCP |
Perplexity and Firecrawl both read the web. Perplexity answers one question. Firecrawl ingests a whole site. Playwright and Chrome MCP both drive a browser. Playwright starts clean, Chrome MCP uses yours. Pick by the job, not the label.
Honest: The Limits You Should Know
None of this is free of trade-offs, and pretending otherwise gets you burned.
Two of the five cost money. Perplexity and Firecrawl need paid API keys, and Firecrawl bills per page, so an unscoped crawl runs up a tab. Set limits before you point them at anything large.
Chrome MCP is a community project, not an official one, and it is the most invasive. It acts inside your logged-in browser with your access. That is the point and the risk in the same sentence. Install it last, and only once you trust how the others behave.
More is not better. Every server you add is one more thing to keep updated and one more surface Claude can act through. Install the two or three that match how you actually work. Leave the rest until you have a real reason.
And these are Claude Code commands. The servers themselves are engine-neutral, so they work in Codex too, but the exact install line differs by tool. If you run Codex, add the same servers through its own MCP config.
If You Only Install One This Week
Install Perplexity. Nothing else on this list.
It is the smallest change with the biggest daily payoff. Live web access under one command, no browser setup, and it fixes the flaw you hit most: Claude answering from stale memory instead of what is true today. Run the one line, ask it to check something current, and see the answer come back from the live web.
Then, once that feels normal, add Playwright. Reading the web and acting on the web are the two capabilities that change what Claude can do for your business, and those two servers cover both. If you want the acting-on-the-web side pushed further, the four-part Claude stack that ships animated sites leans on this same browser control.
Start with one line. If you do not like it, you spent ten minutes. Done.