Connect Claude to Your ERP with MCP: A Practical Guide

You have spent the last half hour opening tabs to figure out how much you have invoiced this month. The number is in your ERP, two clicks away. But those two clicks interrupt whatever you were doing, every single time.

MCP solves that differently from what you are used to: instead of opening the dashboard, you ask Claude. And Claude answers with your real data, not a guess.

This guide shows you how to connect it today, with Frihet, without assuming anything that does not actually work yet.

What you will learn

• What an MCP server is and how it differs from a regular API or a support chatbot.
• What you need ready before you start.
• The steps to connect Claude to Frihet’s MCP server.
• Prompts that already work today, with real examples.
• The limits: what this does and does not do, so you don’t build a workflow on a promise that doesn’t exist yet.

What MCP is, and why it isn’t just another chatbot

MCP (Model Context Protocol) is an open standard for connecting an AI model to external tools and data in a structured way. Anthropic pushed it forward, but it isn’t proprietary: any compatible AI client can speak it.

The difference from a support chatbot is the one that matters: a chatbot tells you where the “new invoice” button lives. MCP gives Claude the same access you have — it can read your real invoicing data and execute the action of creating the invoice, within your account’s permissions.

The difference from pasting data into a generic prompt matters too. When you paste a data export into a chat, Claude sees a fixed snapshot, stripped of context and already outdated the moment you hit send. With MCP, Claude queries the live data, at the moment you ask.

What you need before you start

Three things, none of them complicated:

1. A Frihet account with an active paid plan. API and MCP server access is not included on the free tier — confirm the exact requirement and limits for your plan at docs.frihet.io.
2. An MCP client. The most direct option is Claude Desktop or Claude Code. Both support remote MCP servers natively.
3. Ten unhurried minutes, the first time. The setup is one-time; after that, zero friction.

You don’t need to know how to code. You need to follow one configuration step once, the same way you set up email on your phone once and never touched it again.

Step by step: connect Claude to Frihet’s MCP server

The exact field names, headers, and authentication flow can change between versions of the Claude client. For the literal setup step, the correct source is always the official docs — not a six-month-old screenshot.

If step five works, you’re done. There is no hidden phase two.

For the exact authentication flow and the capabilities exposed version to version, the living reference is docs.frihet.io. This guide gives you the map; the official docs give you the exact coordinates for the day you read them.

Prompts that already work today

This isn’t theory. These are examples of what you can ask or request as soon as the connection is live:

• “How much have I invoiced this month, compared to last month?”
• “Show me the invoices over 30 days past due.”
• “Create an invoice for Estudio Marín SL for 500 euros, brand design, 21% VAT.”
• “Log an 84-euro office supplies expense.”
• “What’s my margin this quarter versus last quarter?”
• “List clients with unpaid invoices, sorted by amount.”

The point: these are questions and actions you currently solve by opening the dashboard, searching, filtering. With MCP, you dictate them to Claude and get an answer in seconds, without navigating anything yourself.

Real use cases

A quick check between meetings. You’re on a call and someone asks how much you’ve billed a client this year. Instead of switching windows, you ask Claude in the chat you already have open.

Frictionless expense logging. Frihet’s OCR already reads receipts and logs expenses from the dashboard. Over MCP, you can ask Claude to log an expense with the details you give it — amount, description, date — without opening the ERP. Photo scanning is still a dashboard feature, not something the MCP server does: for a receipt-to-expense flow, use the app.

A summary before a decision. Before raising prices or taking on a big project, you ask Claude to cross-check your current margin against the last three months of invoicing load. That’s normally a question that means opening two screens and doing the math by hand.

A faster month-end close. You ask for a list of invoices issued, pending, and expenses for the month, in a single answer, to review before handing it to your accountant.

None of these cases require you to build anything. It’s the connection, plus the habit of asking instead of navigating.

Limits and permissions: what this is not

This is where it’s worth being honest, because the noise around “agents that run your business” is inflated.

What you have today: Claude connected to your ERP, running one-off queries and actions that you trigger with a prompt. Every action stays within your account’s permissions and is logged like any other operation in the system.

What you don’t have today, even though it’s the direction the industry is heading: an agent that invoices, collects payment, and reacts to overdue accounts on its own, without you sending the prompt. That’s the agentic economy people talk about — and Frihet is built on MCP and a public API as the foundation for that direction — but it’s a direction, not a box you can check today in settings.

If someone sells you “AI that runs your company on its own” as an already-available product, in any ERP, ask what specific action it executes without a human triggering it. The answer usually clarifies a lot.

Why this is different from a chatbot

A support chatbot lives inside the product and answers with what it knows from the manual. MCP flips that relationship: the AI model lives where you work — Claude Desktop, Claude Code, your editor — and reaches your ERP from there as just another tool.

You don’t switch context to run your business. You run your business from the place you were already thinking about it.

That’s the underlying bet: the less friction there is between thinking of something and doing it, the fewer hours you lose to admin. MCP isn’t just another item on a feature list. It’s the entry point for everything that comes after it.

Common mistakes when starting out

Requesting actions before verifying the connection. If your first request is “create a 3,000-euro invoice” and something in the setup is off, you can end up with a half-completed action. Always start with a read-only query.

Storing the credential in plain text somewhere shared. Just as you wouldn’t paste your bank password into a Slack channel, the MCP credential doesn’t belong in shared notes or public repositories.

Assuming the assistant knows things you haven’t given it permission to see. If Claude answers with generic data instead of your real numbers, the connection isn’t active or the authentication step is missing. It’s not that it “doesn’t work” — you’re still talking to the model without the ERP connected.

Confusing “I can ask it” with “it runs on its own.” You trigger every action. If you stop sending prompts, Frihet doesn’t invoice anything on its own. It’s a tool you use, not an employee working while you sleep — not yet.

The short version

Connecting Claude to your ERP via MCP isn’t a promise for later: it’s a ten-minute setup that already works today, on your paid Frihet account. What is still a future state is the “agent that runs things on its own” part — and it deserves to be treated as exactly that, not as a feature you can already check off a list.

Start by asking. Then, once you trust the answers, start delegating one-off actions. That order is what keeps surprises out of it.