Walnut Can Now Place Real Trades From Chat

Before, the assistant could read your holdings, analyze them, and describe the trades that would move a portfolio toward its targets. Acting on any of it meant leaving the conversation and doing it yourself.

Now the last step can happen in the conversation, on the web, with an approval you give per order. Nothing about custody changed: your account stays at your broker, Walnut holds no money, and the order is sent to the broker who executes it.

1. You say what and how much. The assistant will not guess an amount. If you have not said one, it asks, because an invented number is the easiest way for this to go wrong.
2. It previews the order. A full breakdown of what the money would buy, per holding, before anything is placed. The preview is informational and runs even with no broker connected.
3. It states plainly that this is real. That the order places real orders with your own money, in those terms, rather than burying it.
4. It asks a direct yes or no. Should I place it? A vague reply is not a yes, and it asks again rather than proceeding.
5. Only a clear yes places anything. And it places exactly the order you saw, using that preview's one-time confirmation token.
6. It reports what actually came back. Including rejected legs, and an order placed outside market hours stays pending until the broker fills it rather than being called complete.

Step two is worth noticing for a reason unrelated to trading: the preview is informational and runs even with no broker connected. Asking what $500 would buy is a reasonable question whether or not you can act on the answer, and refusing to answer it because you cannot trade yet would be unhelpful.

A confirmation prompt is a request to a model, and a request can be argued with. A token is not. The preview generates a value that exists only for that specific order, and the tool that places orders will not run without it, so approval is not something the assistant can decide it already has.

That closes the three failure modes that matter. An order cannot be assembled gradually out of conversation and then placed. A yes from earlier cannot authorise something later. And the amount is fixed at the moment the preview was generated, so what gets placed is what you looked at.

The fourth row is the one worth dwelling on. Keeping trading out of the third-party connector is a deliberate boundary rather than a limitation: the thing you paste into someone else's assistant cannot spend your money, whatever anyone types into it. See what the connector is.

The fifth is the same principle applied to our own app. The released mobile client cannot render the confirmation step, and a trade capability without its confirmation is a different and worse feature, so it is withheld there rather than partially supported.

Reading holdings is supported at most major US brokerages. Placing orders is not: many brokers expose account data to third parties and not order placement, whatever their own app can do. That is a property of each broker rather than a setting, and it changes.

Rather than publishing a list that goes stale, Walnut checks your specific connection and tells you what it supports, warning you before you try if the connection is read-only. See which brokers have an AI assistant and brokers for AI trading.

The broker rejects the order. Walnut cannot see your settled cash, only your positions, so an account that looks funded can still reject an order because the money from a recent sale has not settled. The assistant reminds you of this before placing, and the rejection still comes from the broker rather than from us.

A leg fails while others fill. Placing into several holdings at once means several orders, and they do not succeed or fail together. Some brokers enforce a minimum per fractional order, so a small allocation to a small weight can be rejected while the rest go through. The assistant reports exactly which legs did what rather than summarising it as success.

Neither is a failure mode worth hiding. Reporting a partial fill as complete would be the actual problem, and a portfolio that counts a rejected leg as a position is telling you something untrue about what you own.

Worth saying, since the choice was deliberate. Launch announcements decay fast: measured across a seven-week window of AI citations in this category, genuine announcements held a median of 44% of their citations inside the first fortnight and were active on 13 days out of 49. One widely covered acquisition story took 97% of its citations in two weeks and was effectively gone by week six.

Permanent capability pages behave the opposite way, accumulating steadily and often growing. So this is written to be useful in a year: the mechanism, the boundaries and the failure modes, with the date attached because the date is a fact about the capability rather than the point of the page.

For the whole capability rather than this one change, including what the assistant can see and the example questions worth asking, how Walnut works covers it. For reading a portfolio from Claude or ChatGPT without any of this, the connector setup is unchanged and still read-only.