Claude Code MCP Servers: Connecting Claude to Your Business Stack

Introduction

Connecting Claude Code MCP is what transforms Claude from a capable coding assistant into a system that can interact with your actual business tools. For the first six months, I used Claude Code; it was essentially a smart text editor—it could write, reason, and plan, but it couldn’t see my CRM, calendar, booking system, or data warehouse. Every conversation started from scratch because it had no direct access to the information that mattered.

Everything changed when I connected my first MCP server. Suddenly, Claude could read contacts, update my calendar, browse my own websites, and retrieve live data from the systems I use every day. Instead of working with isolated prompts, it could work with real business context.

This guide explains how Claude Code MCP works, why it matters for founder-led businesses, and how to set it up without needing to be an engineer. You’ll learn what MCP actually does, where it delivers the biggest value, and how to connect Claude to the tools that already power your business.

Key Takeaways

• A Claude Code MCP server is a small bridge that lets Claude read and act inside the tools your business already runs on, like CRM, calendar, docs, browser, and design files.
• The MCP standard was published by Anthropic in late 2024 and is now the default way Claude Code talks to external systems.
• Your Claude Code MCP config file lives in a JSON file at the project or user level, and the claude mcp add command writes to it for you.
• The claude code mcp list command shows every server currently connected, including whether it is project-scoped or user-scoped.
• Playwright, Figma, and Context7 are three of the most useful MCP servers for non-engineers because they cover browser, design, and live documentation.
• A small allowlist of MCP servers beats a long one because each server adds tools that the model has to choose between.
• The biggest win for founders is connecting your CRM and calendar, because that turns Claude into a real operator inside your business, not a writing tool.

In late 2024, Anthropic published an open spec called the Model Context Protocol, and within a year, almost every serious tool I touch shipped an MCP server. That is what unlocked the second half of Claude Code for me. The first half is reasoning. The second half is action inside the actual tools you pay for every month.

What a Claude Code MCP Server Actually Is

MCP stands for Model Context Protocol. Strip the name back, and it describes a very simple idea. Claude Code, on its own, is a chat process. It reads text, writes text, and runs commands on your machine. An MCP server is a small program that sits between Claude and some other system, exposing a clean set of tools that Claude can call. The Claude Code MCP layer is the part of Claude that knows how to start those servers, list their tools, and call them with the right arguments.

If you have used a plugin in Word, an extension in Chrome, or an integration inside Zapier, you already understand the shape of it. The MCP server says, “Here are the things I know how to do.” Claude reads that list, and from then on, it can call them like any other tool.

A few examples from my own setup.

• The Google Workspace MCP server lets Claude read my Gmail, draft replies, search Drive, and create Docs.
• The Playwright MCP server lets Claude open a real browser, click around, and screenshot what it sees.
• The Figma MCP server lets Claude read a Figma file, pull out the layout, and turn it into code.
• The Context7 MCP server lets Claude pull up-to-date documentation for any library it is about to use, so it stops guessing at function signatures from training data that is six months old.
• The Nexus MCP server (Nexus is what I call the white-labelled CRM I run my business on) lets Claude pull contacts, look at opportunities, send messages, and update records.

That is the punchline. Claude Code with no MCP servers is a clever colleague who has just walked into your office, knows everything in general and nothing about your business. Claude Code with the right MCP servers is a colleague who can actually do the job because they can see your tools, your data, and your customers.

The Claude Code MCP Config File

Every MCP server you add gets written to a JSON file. The location depends on the scope you chose.

• User scope lives at ~/.claude.json on macOS and Linux, or %UserProfile%\.claude.json on Windows. Servers here are available in every project on the machine.
• Project scope lives at .mcp.json inside the root of the project. Anyone who clones the repo gets the same servers if they trust the file.
• Local scope lives in your project’s local settings and is private to you on this machine.

The format is the same in all three places. Each server has a name, a command to start it, optional arguments, and optional environment variables. A short example.

{
  "mcpServers": {
    "playwright": {
      "command": "npx",
      "args": ["@playwright/mcp@latest"]
    },
    "context7": {
      "command": "npx",
      "args": ["@upstash/context7-mcp@latest"]
    }
  }
}

You almost never have to write this by hand. The claude mcp add command takes a name, a command, and any flags, and writes the entry for you. The reason it is worth knowing where the Claude Code MCP config file lives is for two situations. First, when something goes wrong, you need to see what is actually configured. Second, when you want to share a working setup with a teammate, you can hand them the project-scoped file and they can claude mcp themselves in.

Two practical rules I have settled on.

The first is to keep most servers at the user scope, so I do not have to remember to wire them up in every new workspace. The second is to put any server that holds sensitive context, like my CRM or my business database, into the project scope inside that one repo, so it does not bleed into unrelated work.

The MCP Servers I Actually Run

After about a year of experimenting, I have settled into a small core list. The lesson I keep relearning is that fewer servers, picked deliberately, beat a long list. Every server you add gives the model more tools to choose between, and at some point, that hurts more than it helps.

Playwright for anything browser-based

The Claude Code MCP Playwright server is the one I tell every founder to install first. It hands Claude a real Chromium browser. From inside a session, I can say “open my staging site, click through the booking funnel, and tell me what is broken,” and Claude does it. It clicks, it scrolls, it screenshots, it reads the page. For anyone running a website, a SaaS app, or a marketing funnel, this is the most useful single MCP server I know of.

The killer use case for me is QA. I push a change, ask Claude to walk the funnel, and get a written report with screenshots before I refresh the page myself. It catches broken forms, misaligned mobile layouts, and console errors I would otherwise discover from a customer email.

Figma for turning designs into code

If you work with a designer, the Figma MCP server is the second one to install. The Claude Code MCP Figma integration reads the file directly. It pulls components, layout, type, and spacing and feeds them to Claude when you ask it to build a page. The difference between handing Claude a screenshot and handing it the actual Figma file is enormous. Screenshots make it guess. The file makes it accurate.

I use it most for landing pages and brand updates. My designer pushes a frame to Figma, I point Claude at the URL, and ten minutes later I have a working version of the page inside my own theme.

Context7 for current documentation

This one is small, but it saves me hours every week. The Claude Code MCP Context7 server pulls live, current docs for whatever library or API Claude is about to use. Without it, Claude is working from a model snapshot, which means it sometimes invents functions that no longer exist or misses a flag added six months ago. With Context7, it reads the real docs before it writes the code.

You will know you need it the first time Claude writes a confident block of code that calls a function the library does not have. That is the cue to install Context7 and stop fighting outdated training data.

Google Workspace for the inbox, calendar, and docs

I run Google Workspace, so this is the one MCP server that touches my day-to-day life the most. Claude can search my Gmail, draft a reply to a specific thread, find a document in Drive, or pull a row out of a spreadsheet. Combined with my context files, it lets me say “draft a reply to that prospect who emailed yesterday about pricing,” and Claude knows who I am, what I sell, what I do not say about pricing, and what tone to use.

The Nexus MCP server for CRM

This is the one most founders skip and then regret. Nexus is the white-labelled CRM I run all of my client communications, pipelines, and bookings through. Wiring it up to Claude as an MCP server means Claude can find a contact by name, look at their conversation history, check a pipeline, send a message, or update a tag. Inside a single Claude session, I can ask “what is the status of the deal with that Auckland dental practice we met last week, and draft the follow-up they asked for,” and Claude reads the CRM, checks the conversation, and writes the email in my voice.

Almost every modern business platform now has, or is shipping, an MCP server. CRMs, payment processors, calendar systems, spreadsheet tools, and design platforms are the obvious starting points. If your platform does not have one yet, it is coming because MCP is rapidly becoming the default way Claude and tools talk to each other.

Setting Up Your First Claude Code MCP Server

The shortest path to a working setup is to install one server, get a real win from it, and then add the next one. Trying to install ten on day one is how people get lost.

Step one: pick the first server based on what hurts most

If you push code or run a website, install Playwright first. If you work with a designer, install Figma. If Claude keeps writing code that does not match the current version of a library you use, install Context7. If you spend two hours a day in Gmail, install Google Workspace.

Do not pick the most impressive one. Pick the one that solves a problem you had this week.

Step two: install with claude mcp add

In your terminal, inside any project, run the command for your chosen server. Most servers have a one-line install in their README. For Playwright, it looks like this.

claude mcp add playwright npx @playwright/mcp@latest

For a stdio server with environment variables, you pass them with -e flags.

claude mcp add my-crm \
  -e API_KEY=xxxx \
  -e BASE_URL=https://api.example.com \
  -- npx some-crm-mcp@latest

The -- separates Claude’s own flags from the command that starts the server. If you forget it, the env flags get attached to the wrong process, and the server fails silently.

Step three: confirm with Claude Code MCP list

Run claude mcp list and you will see the server listed. The output also tells you whether it is project-scoped or user-scoped, and whether Claude can currently connect to it. The Claude Code MCP list command is the first thing I run when something is not working, because half the time the answer is that a server is at the wrong scope or has not started.

If you see your server listed but marked as failed, restart your Claude Code session. The server only starts when Claude tries to use it, and any change to the config needs a fresh session to load.

Step four: ask Claude to use it

This is where most people get stuck. They install the server, see it in the list, and then forget that they need to actually invoke its tools. Open a Claude Code session and ask a question that requires the server. For Playwright, “open https://octavius.ai and tell me what you see on the homepage.” For Figma, “read this Figma file and tell me what components are on the hero section.” For Context7, “use Context7 to pull the latest Anthropic SDK docs and confirm the prompt caching syntax.”

The model will choose to use the new tool, and you will see the tool calls in the transcript. That is the moment it clicks. Claude is not just talking about your tools anymore. It is using them.

Step five: lock down what each server can do

Some MCP servers expose dangerous tools by default. A CRM server might let Claude delete contacts. A file server might let it overwrite a database. Read what each server can do before you give it permission to act without asking. The Claude Code permission system lets you allowlist specific tools per server, and I strongly recommend doing this for anything that touches paid systems or production data.

What Happens Once Your Stack Is Connected

Most founders I talk to think the MCP layer is a technical detail. It is not. It is the bridge between “I have an AI that can think” and “I have an AI that can do the job.” Until the bridge exists, the AI is a clever advisor sitting in a meeting room with no laptop, no phone, and no access. Once the bridge exists, the AI is an operator with a desk inside your business.

That changes what is possible. With a connected stack, I can.

• Ask Claude in the morning what changed in my pipeline overnight, and get a real answer drawn from the CRM, not a guess.
• Push a draft email to a specific Gmail thread without leaving the terminal.
• QA a new landing page in a real browser before I tell my team it is ready.
• Pull up current docs for any library while building, instead of trusting Claude’s training cut-off.
• Draft a new Figma layout into working code in one pass.

None of these is theoretical. They are how I run my own week. The pieces that make it possible are the MCP servers, plus a small set of structured context files that teach Claude who I am, what I sell, and how I work.

That second piece is the part most people miss. A connected stack is necessary but not sufficient. Without context, Claude can read your CRM but does not know which deals matter. Without context, it can write code but does not know your brand voice. The MCP layer gives Claude reach. The context layer gives it judgment. Together, they start to behave like an actual member of the team.

This is what I now build for clients in my work at Octavius, and it is also what I built for my own business first. If you want to see how the connected layer plus the context layer come together into an AI brain plus AI workforce setup that actually runs alongside the founder, I have written about the AI brain and the system that runs your business and about why most AI initiatives stall in two earlier posts.

Common Pitfalls and What to Do About Them

A short list of the mistakes I have made or watched other founders make.

Installing too many servers at once. Every MCP server adds tools that the model has to read at the start of every session. Past about eight to ten servers, the model gets noticeably slower and worse at picking the right one. Start with two or three. Add more only when the lack of one is causing real pain.

Forgetting to restart the session after a config change. Claude reads the MCP config when a session starts. If you add a server mid-session and it is not appearing, exit and reopen. This trips up almost everyone the first few times.

Leaving secrets in the config file. Treat your MCP config the same way you would treat your .env file. Most servers accept environment variables, and you should pass API keys through env, not hard-code them into JSON.

Not testing what each server can do. Read the README for each MCP server before you install it. Some servers can write, delete, send messages, or charge cards. You want to know exactly what is exposed before you turn it on.

Installing the wrong scope. A user-scope server appears in every project. A project-scope server is locked to one repo. If you mix them up, you will either be surprised that a CRM server is showing up in an unrelated workspace or confused that your shared server is missing in a new project.

Treating MCP as the whole story. The connected stack is half the answer. The other half is the context files that teach Claude your business. A founder with rich context files and three MCP servers gets more out of Claude than one with thin context and twenty servers.

The Bigger Picture: From Connected Tools to a Real System

Once you have a few MCP servers wired up, you stop thinking about “AI tools” and start thinking about “the system.” That shift in language is the shift that matters. The system is your context files, your data layer, the morning briefings Claude writes for you, and the MCP servers that let it act inside your actual business.

This is the working pattern I see when a founder makes the jump from “I use ChatGPT sometimes” to “the business is starting to run itself a bit.” The MCP layer is what makes the jump physically possible. Before MCP, you needed an engineer to wire each integration by hand. After MCP, you run one command, paste an API key, and Claude is connected.

The market is moving very fast right now. I expect that in twelve months, every serious business tool will ship an MCP server alongside its public API, and the friction to connect Claude to your stack will be close to zero. Founders who get familiar with this layer now will be far ahead of the ones still pasting screenshots into ChatGPT in a year’s time.

Conclusion

The biggest shift happens when this technology stops being a technical feature and becomes a fundamental part of your daily workflow. On its own, Claude can reason, write, and plan. Once it’s connected to the tools your business already relies on, it can act on that reasoning using real data and live systems.

Integrating Claude Code MCP allows you to turn useful conversations into completed work by giving the AI direct access to your internal processes. Each new connection increases what Claude can accomplish without requiring you to repeat context or manually move information between platforms.

The best approach is to start small by connecting one service that solves a genuine bottleneck. Over time, those integrations compound into a connected operating environment where AI isn’t just generating ideas—it’s helping your business execute them.

Take the Next Step

If you can see the shape of this for your own business and want a hand mapping it out, book a free 30-minute Discovery Call here: Octavius Discovery Call. It is a short conversation where I ask about your stack, the work you are stuck doing yourself, and where Claude Code plus a connected set of MCP servers could realistically take a load off. No pitch deck, no slides, just a quick read on whether this is the right next step for you.

Frequently Asked Questions

About Octavius

Titus Mulquiney is the founder of Octavius AI, where he builds AI brains and AI workforces for founder-led businesses stuck running everything out of their own head. Twenty years in marketing, ex-Sony product manager, ex-GM Zeal NZ. Based in Auckland, working with operators across NZ, Australia, and the US. Connect on LinkedIn.