Blog · MCP & agents

What Is an MCP Server? Explained for Non-Developers

An MCP server gives your AI assistant hands: real tools it can use in outside software. What the Model Context Protocol is, how connecting works, and what to check before you trust one.

· 7 min read · by the LedgerMCP team

An MCP server gives your AI assistant hands. Out of the box, Claude and ChatGPT can only talk; connect them to an MCP server and they can do: look things up in real software, and take real actions in it. MCP (the Model Context Protocol) is the open standard that makes this work the same way everywhere, and it's the reason a chatbot can now keep your books, manage your calendar, or search your company's documents. Here's the whole idea, no code required.

The problem MCP solves

Before MCP, every AI-to-software connection was a custom, one-off integration. The AI vendor had to build support for each app, and most apps never made the list. So your assistant could write beautifully about bookkeeping while being unable to touch your actual books. Anthropic published MCP as an open standard in late 2024, and it has since become the common language: any assistant that speaks MCP can use any tool that serves it. One plug shape, every appliance.

The three pieces, in restaurant terms

  • The MCP client is your assistant: Claude, ChatGPT, or others. It's the customer who orders.
  • The MCP server is the kitchen of a specific piece of software. It publishes a menu of tools, meaning named actions the software is willing to perform.
  • Tools are the dishes: precise, named capabilities like import_transactions or run_pnl. The assistant orders them; the software executes them.

When you tell your connected assistant “categorize June,” it reads the menu, picks the right tools, fills in the parameters, and places the order. The software does the actual work; the AI never reaches past the menu. That boundary is the point: the server decides what's possible; the AI only decides what to ask for.

What connecting looks like (it's genuinely one minute)

Modern clients connect to a hosted MCP server with a URL and a sign-in. In Claude: Settings → Connectors → add the server's address. In ChatGPT: Settings → Apps (developer mode) → add by URL. You approve access by signing in (for LedgerMCP's server that's a 6-digit email code), and from then on the assistant has the tools in every conversation. Power users and scripts can use an API key instead; the how-it-works page shows all three paths.

“Should I trust an MCP server?” The right questions

Connecting a server means letting your assistant act in that software as you. So evaluate the server the way you'd evaluate an employee's permissions:

  1. Who runs it? A hosted server from the software vendor itself beats a third-party wrapper holding your credentials. (This matters in accounting specifically: most “QuickBooks MCP servers” are community projects you must trust with your books' keys.)
  2. What can its tools actually do? Read the tool list; a good server publishes it. Is there a read-only mode for cautious starts?
  3. What happens when the AI errs? The best servers are built so mistakes are contained: actions logged, reversible, dangerous operations simply not offered as tools.
  4. Can you revoke access instantly? Disconnecting or revoking a key should take one click.

This is where server design does the heavy lifting. LedgerMCP's server, for example, offers 52 bookkeeping tools but no delete tools. Underneath, the database refuses unbalanced entries and keeps history immutable, so even a malfunctioning assistant can only make reversible, logged changes. The full safety argument is here.

Why this beats “AI features” built into apps

An AI feature inside an app is that vendor's model, on that vendor's terms, inside that vendor's walls. MCP flips the ownership: your assistant, the one whose subscription you pay and whose behavior you know, works across every tool you've connected, and can coordinate between them in one conversation. Switch assistants next year and your tools come along; compare Claude vs ChatGPT freely, because the server can't tell the difference.

MCP in one worked example

You attach a bank statement and say “import this and close out June.” Your assistant calls import_transactions (84 rows land, deduplicated), bulk_categorize (81 posted, 3 flagged for you), confirm_transfer (a move between your accounts posts once, not as fake income), reconcile_account (ties to the statement, to the cent), and set_lock_date (June is now closed). Every call is logged; any can be reversed. That's an MCP server doing its job: the assistant supplied judgment and conversation, the software supplied capability and guardrails.

Curious to try one? Bookkeeping is about the best first MCP experience there is: the work is real, the value is immediate, and the guardrails are strict. Start with the complete AI bookkeeping guide, or just connect Claude or ChatGPT to a free set of books and hand it a statement.

Put this into practice

Free books in one minute: connect Claude or ChatGPT and let it do the work you just read about.