Can ChatGPT See Your Portfolio? What ChatGPT Finances Can and Can't Do

Finances in ChatGPT lets US users link financial accounts through Plaid: banks, credit cards, brokerages and loans, plus credit report data through Experian. It is available on web, iOS and Android, on the Free, Go, Plus and Pro plans. Once an account is linked, ChatGPT can show:

• Balances, transactions and net worth across linked accounts.
• Investments, including a market update with daily changes, current value and position size for equity holdings.
• Stock watchlists.

Two caveats. Some institutions share balances but not holdings, so a linked brokerage can appear as a single number without its positions. And if you do not link anything, or you live outside the US, ChatGPT is back to the old situation: it knows only what you paste into the chat or what it can search.

OpenAI is explicit about the limits. Finances in ChatGPT cannot move money, pay bills, make trades, open or close accounts, or file taxes, and it is not an adviser. For investing, three gaps matter most:

• No trades. It can describe a position. It cannot buy or sell one, so any change you decide on happens back in your broker's own app.
• No target weights. It reads what you hold. It has no notion of what you meant to hold, so there is no basket, no target allocation, and no rebalance toward one.
• US only. Investors in Canada, the UK, Australia and elsewhere get nothing from Finances.

Walnut is an AI investing assistant built around acting on a portfolio, not only reading one. It connects your existing brokerage securely (read-only by default) and adds what Finances leaves out:

• Trades you approve. At brokers that support it (E*Trade, moomoo, Public, Questrade, tastytrade, Webull and others), Walnut can place orders, and only with your explicit approval each time.
• Thematic portfolios with target weights. Baskets hold a stated thesis and target weights, and Walnut can suggest the trades that would bring a basket back to them.
• Works in Claude too. The same connection is available inside Claude, ChatGPT, or Walnut's built-in assistant, and it is not limited to US brokers.
• Each holding framed against the S&P 500. Performance is scored against the index, not only reported as a value.

Where ChatGPT wins: it is built in, free to link on every plan, and covers banks, cards and loans as well as investments, which Walnut does not.

The more surprising failure is not that ChatGPT says “I don’t know.” It is that it answers with a specific figure that sounds right and is wrong. This happens because a language model is built to produce a plausible continuation. When you ask for a price, a return, or a position size it has no live data for, it generates a number that fits the pattern of the question rather than refusing.

The tone is the trap. The confident phrasing is identical whether the figure is grounded or invented, so there is no built-in signal telling you which is which. This is why the safe habit with any general model is to treat every specific number as a claim to verify, not a fact to act on, until it is grounded in real data from your broker or a live source.

The honest picture changed in 2026. ChatGPT was always strong at explaining and reasoning; with Finances it can now read linked US accounts too. What it still lacks is the ability to act. This table draws that line.

The pattern: reading and explaining, it can do (in the US, once linked). Acting on your account, by trading or working toward target weights, it cannot.

If Finances is not available to you, the simplest workaround needs no new tool. Paste your holdings into the chat: tickers, share counts, and current values, and ChatGPT can reason over exactly what you gave it. For a one-off question, this works well.

The limits are real, though. It only knows what you typed, so it misses anything you left out. The snapshot goes stale the moment prices move. You have to redo it every session, and copying live values by hand is error-prone. It is a fine patch for a single conversation, but it is not a standing view of your portfolio. For a walk-through of doing this well, see how to use ChatGPT to analyze your portfolio.

To act on your holdings, or to use your real portfolio in Claude or outside the US, you connect your accounts through a tool that sits between your broker and the model. The model never talks to your broker directly; a connector or aggregator does, and controls what data is exposed and what actions are allowed. There are two common paths, and neither is the only right answer:

• A general MCP connector. MCP (Model Context Protocol) is an open standard that lets an AI assistant call external tools and data sources. An MCP connector for a brokerage exposes your account data in a structured way, so an assistant can reason over your real holdings instead of guessing. For the general approach, see how to connect a brokerage to an AI assistant.
• A connected assistant like Walnut. Walnut is an AI investing assistant you chat with over the broker you already own. It connects your existing brokerage (read-only by default) and lets you ask about what you actually hold through Claude, ChatGPT, or a built-in assistant, with each holding framed against the S&P 500. You approve every trade, and Walnut is not an investment adviser.

Both approaches solve the same core problem: they put your real data in front of the model, so the conversation is about your account instead of an abstraction. Which you choose depends on how much you want to assemble yourself versus use a ready-made connected product. Walnut is one concrete option among several.

The gap is not only that a pasted list goes stale. It is that some of the most useful things about a portfolio are arithmetic over the whole set of weights, and that arithmetic is only as good as the weights you supplied. People paste tickers far more often than they paste accurate current market values, so the answer is computed on a portfolio that does not exist.

Concentration is the clearest case. The question is not how many positions you own but how many the portfolio actually behaves like, and the two diverge as soon as your weights stop being even. Ten holdings at equal weight behave like ten. Ten where one has grown to 40% behave like about 4.4. Getting that right requires every position and its current value, not a list of names, and a general assistant working from what you typed has neither.

Walnut computes this as Effective Holdings from the account you connect, and publishes the formula so you can check it in a spreadsheet. It is deliberately a description rather than a rating: there is no correct level of concentration, and the useful signal is usually the gap between the number you own and the number you behave like, which is almost always a surprise.

This is also why the answer changes on its own. A pasted portfolio is a photograph, and concentration drifts continuously as prices move, so the position that was 22% of your account last year may be 38% now without you having bought a single additional share. A connected tool recomputes it; a photograph cannot.

Whichever path you take, a few filters keep it safe and useful:

• Read-only by default. Prefer a connection that reads your holdings without the standing ability to move money. Walnut is read-only by default and connects through a regulated aggregator.
• Approval for any action. If a tool can place trades, it should require your explicit approval each time, not act on its own.
• Grounded, not guessed. The point of connecting is real data. Make sure the tool actually feeds your live holdings and prices to the model rather than paraphrasing them.
• Clear on advice. A trustworthy connected tool explains and frames trade-offs without pretending to be your adviser. Be wary of anything promising guaranteed market-beating returns.
• Credentials stay out of the chat. Never paste broker passwords into a raw chat. Connections should run through a proper aggregator or OAuth flow, not a copied login.

Most people try two or three of these before concluding the problem is structural. It is worth knowing why each fails, because each fails differently.

• Pasting a holdings list. Works once, for questions about what you typed. It goes stale immediately, and people paste tickers far more often than current market values, so anything that depends on weights is computed on a portfolio that does not exist.
• Uploading a brokerage statement. Better, because the weights are real, and still a photograph. It is also the version where people most often hand over more personal data than the question needed, including account numbers that no analysis requires.
• Screenshots. The model reads what is legible in the image, which is usually the top few holdings, so the answer silently describes part of your portfolio as though it were all of it. This is the most misleading option because nothing announces the truncation.
• Asking it to browse your broker. It cannot. Your account sits behind authentication, and browsing has no credential to reach it. Anything that appears to work here is reasoning from public information about the ticker. In the US, linking through Finances is the route that actually reads the account.

The common thread is that all four move data to the model by hand, once. What changes the situation is a connection the model can read on demand, such as Finances in the US or a connected tool, which is a different mechanism rather than a better version of the same one.

Connecting an account is the fix, and it is worth doing deliberately rather than clicking through. Four questions, in order of how much they matter.

• Is the access read-only, or can it trade? A read-only connection cannot place an order at all, because the access it holds does not carry the permission. Your broker shows which one you granted in its connected-apps settings, and if the broker and the tool disagree, believe the broker.
• Where does the money live? A tool that reads your existing account never holds your money. A service that manages a portfolio opens an account and holds it. These are different products and the distinction is easy to miss during onboarding.
• Who is the aggregator? Most tools do not build broker connections themselves; they use a regulated data aggregator. Knowing which one tells you whose security practices you are actually relying on.
• How do you revoke it? Find the control before you need it. Revocation exists at your broker as well as in the tool, and the broker's version is the one that actually cuts the connection.

None of this makes connecting risk-free. Read-only access still exposes what you own, so you are trusting the provider with your holdings even when they cannot act on them. It is a smaller blast radius, not zero, and connecting only what you need is the sensible default.

In the US, ChatGPT can now see your portfolio once you link it through Finances, on every plan including Free. It still cannot place trades, hold a portfolio to target weights, or help anyone outside the US, some institutions share balances without holdings, and a model can still state a wrong figure with confidence. If you want to act on your holdings, a connected tool covers that ground. Walnut is one option, not the only one: it connects your brokerage securely, keeps access read-only by default, places trades only with your approval, holds thematic portfolios to target weights, works in Claude as well as ChatGPT, and frames each holding against the S&P 500. Walnut is not an investment adviser.

For the wider field of tools, see the best AI investing tools roundup.