← Field notes// field note

Nmbrs MCP: connect Nmbrs Accounting to ChatGPT, read-only

Your bookkeeping already knows the answer. With an MCP server in between, you can ask ChatGPT instead of clicking through 246 administrations.

October 2026

6 min read

An Nmbrs MCP server is a small connector between ChatGPT (or Claude) and the Nmbrs Accounting API, the old Yuki web services. ChatGPT asks for balances, ledger entries, open items or archived documents, the server fetches them read-only and returns text. Nothing is written back, the server makes no AI calls, and it runs in the EU.

Nmbrs MCP: ChatGPT connected read-only to Nmbrs Accounting, QuikCode field note

You can connect Nmbrs Accounting to ChatGPT with an Nmbrs MCP server: a small connector that sits between ChatGPT (or Claude) and the Nmbrs Accounting API. You ask a question in plain language, ChatGPT picks the right tool, the connector fetches the data from Nmbrs read-only and hands back text. No exports, no copying numbers into a chat.

I built this for Landsman en Teune, a bookkeeping and advice office in Kampen with around 246 active client administrations in Nmbrs Accounting (the product that used to be called Yuki). They already worked in ChatGPT Business. What they missed was a way to let it look things up in their own systems without uploading files by hand. The full story is in the Landsman en Teune case. This note is about the Nmbrs part.

What MCP is, in one paragraph

MCP (Model Context Protocol) is an open standard for giving an AI assistant tools. The connector describes what it can do ("find an administration", "get ledger balances on a date"), and the assistant decides when to call which tool. ChatGPT Business supports custom MCP connectors, and so does Claude. So the same server works with both. The AI stays in the client's own ChatGPT account, under their own contract with OpenAI. My server never calls an AI model itself: it fetches, converts to text, and returns.

What ChatGPT can read from Nmbrs

These are the tools the connector offers. All of them read, none of them write.

That turns into questions like "which debtors in this administration have been open for more than 60 days?", "find the lease contract for this client and tell me when it ends" or "what was the balance on the bank account on 31 December?". ChatGPT chains the tools itself: find the administration, pull the entries, open the document behind the odd one.

How the Nmbrs Accounting API actually works

If you plan to build this yourself, these are the things I'd want to know up front.

  1. It's SOAP, not REST. Nmbrs Accounting still runs on the Yuki web services. You authenticate with an access key, get a session, and send XML envelopes. Not hard, just old-fashioned.
  2. Switch domain before every question. One access key can see many administrations, but each query runs against the "current domain". Forget to switch and you get an answer about the wrong client. The connector sets the right domain for every single call, so that mistake can't happen.
  3. There is a daily call limit. It depends on your package. This client started at 1,000 calls a day and later raised it to 5,000. An AI assistant that chains tools burns calls quickly, so the connector caches answers for an hour and logs a warning well before the limit.
  4. Some things are simply not in the API. The quality monitor, the tasks and deadlines from the portal, and the VAT code per individual entry are not available. I'd rather tell a client that on day one than discover it in week three. Also: an empty workflow doesn't prove a period is closed. The tool says so in its own description, so ChatGPT doesn't claim otherwise.

Why read-only, and what the logs keep

Bookkeeping data is client data, so the limits are built into the software, not into a policy document.

I wrote more about this side in is ChatGPT GDPR-safe with client data.

What you can connect next to Nmbrs

Nmbrs was one of three systems behind the same connector. The others were Accelerate (the advice software) and two Outlook mailboxes, where ChatGPT can search and prepare drafts but cannot send. That is the real value: one question that combines the ledger, the mailbox and the client file.

The same pattern works for most systems with an API:

You can see more builds like this on my work page.

When this is not the right approach

Most of the work is in the boundaries: which data, for whom, read or write, and what gets logged. If you want your bookkeeping to answer questions in ChatGPT, let's talk for 30 minutes.

Same question for Twinfield, AFAS or SnelStart? Here is how I connect Claude or ChatGPT to your accounting software, package by package.

Is there an MCP server for Nmbrs?

You can build one on top of the Nmbrs Accounting API, which is what I did for a bookkeeping office. It exposes read tools such as balances, ledger entries, open items and archived documents, and ChatGPT or Claude call them when you ask a question.

Can I connect Nmbrs Accounting (Yuki) to ChatGPT?

Yes, with a custom connector. ChatGPT Business supports custom MCP connectors. The connector authenticates against the Nmbrs Accounting web services, switches to the right administration for every call and returns the data as text to ChatGPT.

Does the Nmbrs Accounting API have a limit?

Yes, a daily call limit that depends on your package. In this build it started at 1,000 calls a day and was raised to 5,000. Caching answers for an hour keeps an AI assistant comfortably under it.

Is it safe to give ChatGPT access to bookkeeping data?

It can be, if the limits are in the software: read-only tools, personal logins, logs without content, EU hosting and a data processing agreement. The AI runs in your own ChatGPT Business account, under your own contract with OpenAI.

Is our bookkeeping data used to train ChatGPT?

No. The AI runs in your own ChatGPT Business account, where data isn't used for training, and the connector itself makes no AI calls. It only fetches data and returns text.

Does this also work with Claude?

Yes. MCP is an open standard, and Claude supports custom connectors as well. The same server can serve both, so the choice of assistant stays with you.

Got a process that should be software?

Book a call
© 2026 QuikCode · software for better operations