Ask your portfolio: Valued's read-only connector for ChatGPT

Since today, 17 September 2026, a Valued account can be connected to ChatGPT. The assistant can then read the workspace you select (investments, and if you allow it stored document text, e-mails and notes) and answer questions about it with a link to the page in Valued each answer came from. It can only read. There is no tool in the connection that changes an investment, uploads a document, deletes anything or sends an e-mail.

This article says what the connector reads, what it leaves out, how access is granted and ended, and why we stopped at reading.

What the connector is

The connector is a small server of ours that speaks the Model Context Protocol (MCP), an open standard for connecting AI applications to external systems. An assistant that supports MCP can call the tools such a server offers. Ours offers four, and all four read: list the investments of the workspace, list the records of one kind (capital calls, distributions, cap table rows, loans, transactions, valuations and so on), search by a word or phrase, and fetch one source in full.

The first client we support is ChatGPT. For now the connection is added by hand: you switch on developer mode in ChatGPT's settings, add an app with our server address and choose OAuth as the sign-in. Whether that option is offered depends on your ChatGPT plan and, in an organisation, on its policy. The steps are under Settings, in the ChatGPT tab of your Valued account, and take about two minutes.

What it can read

When you connect, ChatGPT opens Valued, you sign in there, and a consent page asks two things: which workspace, and which of four categories.

Investments and financial records cover the four areas of the product: fund commitments with their capital calls, distributions and quarterly report figures; equity investments with cap table rows, loans and valuations; depots with holdings and transactions; properties with loans, leases and cash flows. The other three categories are stored document text and summaries, stored e-mails, and investment notes.

Nothing is preselected. A connection covers exactly one workspace, and a category you leave unticked stays closed even if the assistant asks for a record by its identifier. If you connect with investments only and later ask about e-mails, the answer you should get is that this connection was not given access to e-mails, not that there are none. We had to teach the tool descriptions that difference during testing, because an assistant that finds nothing will otherwise say there is nothing.

With that in place the questions it handles well are the ones that are tedious to click together: which capital calls are still open and when they are due; what the last quarterly report of a fund said about NAV; which distributions have been announced and not yet received; who holds what on a company's cap table. You can also give ChatGPT a file of your own and ask it to compare the file with what Valued has stored. The file stays on ChatGPT's side.

What it cannot do, and what it will not pretend

It cannot write. That is not a setting. The server has no writing tool, each database call runs in a read-only transaction, and the database role it uses has no direct access to any table, only to one function that checks the connection on every read and returns rows of the connected workspace.

It does not hand over original files. For a document, the connector returns the text or summary Valued already holds, and it says so: a source that exists only as a summary is marked as such, a long one as truncated. An honest answer therefore sometimes reads "according to the stored summary", and nobody should take that for a reading of the PDF.

It does not calculate for you. The connector returns stored records and stored figures; it offers no portfolio total and no freshly computed IRR. Amounts come in their own currency, and the tool description tells the assistant not to add euros to dollars. If you ask for the total value of a portfolio held in two currencies, a sum per currency or a question back is the right answer. A single number would contain an exchange rate the assistant made up.

And the answers are ChatGPT's. The connector supplies records; the sentences around them are written by a language model, which can misread a question or stop paging through a long list too early. The link to the page in Valued is there so that you can check, and for anything you act on, you should.

How access is granted and how it ends

Access follows your Valued login, including two-factor authentication if you use it. Every new connection is announced to you by e-mail. It then ends in one of five ways: you disconnect it under Settings, in the ChatGPT tab; 30 days pass without use; 90 days pass, at the latest; you change your password or your two-factor settings; or you leave the workspace. Signing out of Valued in the browser does not end it, which is deliberate: the connection is meant to work while you are not logged in.

One thing disconnecting cannot do is reach back. What ChatGPT has already read stays in that conversation, on OpenAI's side and under OpenAI's terms, not ours. This is the real decision you make when you connect, and it is the subject of our June article on AI assistants and your financial data. Check the data controls of the assistant before you give it a portfolio, and share only the categories the question needs. Our side of the record is short. We log that a read happened, by which connection and for which record, and not the search text or the content returned.

Why read-only

Two reasons, and the second matters more.

The first is the obvious one. An assistant that can book a capital call can book a wrong one, and a platform's own rules for such apps already exclude the dangerous end of this: OpenAI's guidelines prohibit apps from executing money transfers or investment trades.

The second is specific to how Valued works. Everything in a Valued portfolio got there because a document was read, the figures were checked, and you approved them. We wrote in February about what a model reliably reads in a capital call and where it fails. An assistant with write access would be a second way in that skips the approval. It would also be exposed to the content it reads: a stored e-mail or PDF can contain text that looks like an instruction, and a model that reads it may follow it. With no writing tool, the worst such text can do is produce a wrong answer. It cannot produce a wrong booking.

So questions go through the connector, and changes go through the product, where you see the document next to the figure before you confirm it.

Also new this month

Two other things arrived in September. Valued now has a German interface, so the product can be used in German as well as English. And real estate is an area of its own, next to funds, equity investments and securities: properties with their loans and fixed-rate periods, rent roll, income and expenses, and yields.

If you already use Valued, the place to start is Settings, then the ChatGPT tab. Connect one workspace with investments only, ask which capital calls are open, and compare the answer with the fund overview you know.