FacebookTwitterLinkedInTelegramCopy LinkEmail
Blockchain

Coinbase Brings User-Approved Trading Into ChatGPT

Coinbase Brings User-Approved Trading Into ChatGPT

Eligible Coinbase customers in the U.S. can now place crypto, stock and futures orders from inside ChatGPT, but every trade still waits on a tap from the account holder.

Key Takeaways

  • Every plugin order needs a confirmation card.
  • A confirmed order can still sit unfilled.
  • Agent futures trading uses your main portfolio.
  • Auto-stop pauses trading; it doesn’t sell holdings.

Coinbase announced the official plugin on November 17, five months after launching Coinbase for Agents on June 11, the platform that first let AI assistants connect to Coinbase accounts.

Once an account is linked, a user can view the connected portfolio and ask ChatGPT to place, edit or cancel orders. Each of those actions produces a Coinbase confirmation card, and nothing goes to the exchange until the user approves it. The plugin is limited to U.S. customers. Which products a user can trade depends on their Coinbase account eligibility, and access also depends on their ChatGPT account and workspace.

Orders are routed through Coinbase Advanced, and Coinbase’s developer documentation says Advanced Trade fees and eligibility requirements apply. A trade requested in a chat is priced under the same fee schedule as one entered on the exchange directly.

The plugin is not the only way to connect the two services, and the route a user picks decides who makes the final call on each order.

Two ways to put Coinbase inside ChatGPT

Official Coinbase plugin

Setup: Coinbase’s own plugin, linked to the account

Before an order leaves: A confirmation card for every order, edit and cancellation

x402 payments: Not offered

Custom MCP connection

Setup: Added by hand as a custom MCP server; may require ChatGPT developer mode

Before an order leaves: Whatever the user instructs; in Coinbase’s tests, ChatGPT placed clear orders in a single turn

x402 payments: Available

MCP, short for Model Context Protocol, is the standard AI assistants use to call outside tools. x402 is Coinbase’s protocol for software paying for data or services, typically in stablecoins.

Both routes run in the same chat window and describe orders in the same fluent language. What separates them is whether a person approves each order, or whether the original instruction already counts as approval.

The card shows what ChatGPT understood

Take a plain request: “Buy $100 of Bitcoin with a limit order.” It names the asset, the direction, the amount and the order type. It does not say the highest price the user will pay, how long the order should stay open, whether the $100 includes fees or which portfolio should fund it.

Those points have to be settled before a card can appear, either by ChatGPT asking or by defaults filling the gaps. Coinbase’s documentation says assistants reliably build orders “from clear, well-specified instructions”, and the vaguer the request, the more of the order the assistant decides. The lines worth reading on the card are therefore the ones the user never typed: the limit price and the duration.

Approval starts an order; it doesn’t finish one

Tapping approve sends an instruction. Whether anything was bought depends on what happens after Coinbase receives it, and a chat reply can run ahead of the facts. Coinbase’s agent documentation warns that on hosted chat clients, high-reasoning modes have sometimes drafted an order and stopped before executing it, leaving the user to ask again.

An order that does reach the exchange can still sit unfilled. Because the plugin covers stocks as well as crypto, equity rules come with it. Coinbase’s stock-order guidance says limit orders execute only at the set price or better when enough liquidity is available, so they can fill partially or not at all. Stock market orders are limited to regular trading hours, and extended sessions carry their own restrictions.

Partial fills change what a cancellation does. Cancelling a $100 buy after $60 has executed removes the remaining $40. The $60 position stays in the account at whatever price it filled.

When the instruction is the approval

On the custom connection, an agent can act on standing instructions instead of waiting for a tap on each order. Coindoo has examined how far automated execution sits from independent AI decision-making. For a Coinbase user, the more immediate question is which money the agent can reach.

Coinbase’s answer is portfolio scope. Its documentation advises creating a separate portfolio for the agent, funding it only with what the user is prepared to risk and selecting that portfolio on the approval screen. Transfers through the agent connection are limited to moves between the user’s own portfolios, with no withdrawals to outside addresses.

That isolation covers spot crypto only. Coinbase’s U.S. dated futures work only in the default portfolio, so a user who wants an agent to trade them has to add the default portfolio to the agent’s access, and with it everything else that portfolio holds.

Coinbase’s server-side limits and what they leave open

Coinbase added another layer on November 11. Its Agent Controls announcement lets users cap the dollar size of each trade, cap an agent’s total trading over 24 hours and set an auto-stop that pauses the agent once a portfolio falls by a chosen percentage. The limits are enforced on Coinbase’s servers and sit behind the user’s login, so the agent cannot loosen them. Permissions by product type were marked as coming soon.

The two caps stop different failures. A $100 per-trade limit blocks one oversized order but still allows fifty $100 orders in a day. The 24-hour cap is what stops the fiftieth, because trades beyond it are not executed.

Auto-stop is narrower than its name suggests. Coinbase describes it as halting new agent trades until the user switches trading back on. It has not described it as cancelling orders already resting on the book or closing positions, so anything bought before the trigger keeps moving with the market after the agent goes quiet.

A conventional stop order actively tries to exit, but it carries a gap of its own. Coinbase’s U.S. derivatives guidance explains that when a market reopens beyond a stop level, the order can fill at the opening price rather than at the trigger.

Agent Controls arrived six days before the plugin and are described in terms of agents acting on a user’s behalf. For trades placed through the plugin, the control Coinbase has put in front of every order is the confirmation card itself.

Checking the result in Coinbase

After approval, the record that counts sits in Coinbase’s order history, not in ChatGPT’s reply. It shows the order’s status, how much filled, the average price, the fees charged and any quantity still open. A partly filled limit order left open is still a live instruction, and a futures position an agent opened stays in the default portfolio after any auto-stop.

Neither the plugin nor Agent Controls closes a position on the user’s behalf. That still takes an order of its own, approved like every other.

Disclaimer: This article is for informational purposes only and does not constitute financial, investment or legal advice. It describes Coinbase’s ChatGPT plugin and agent features as publicly documented on November 17, 2026; availability, supported products, fees and controls may change and depend on individual account eligibility. Trading cryptocurrencies, stocks and futures involves substantial risk, including the loss of more than an initial position in leveraged products. Check Coinbase’s current terms and documentation before connecting any AI tool to a trading account. 

Author

Reporter at Coindoo

Alexander Zdravkov is a market analyst and crypto journalist with interests in economics, broader financial markets and digital assets. His journey into crypto began more than four years ago, driven by a fascination with the rapid evolution of blockchain technology and the transformative potential of decentralized finance. He began analyzing market cycles and identifying emerging trends before they reach the mainstream. He holds a degree in International Relations - a background that helped shape his broader perspective on global economics, geopolitics, and the interconnected nature of modern financial markets. Whether covering the latest developments in the crypto sector or exploring broader macroeconomic themes, Alexander focuses on giving readers context rather than simply repeating headlines. During his career, he has authored more than 5,000 articles covering cryptocurrencies, traditional finance, and global market developments. His work spans everything from Bitcoin and altcoins to macroeconomic trends influencing risk assets worldwide.

Learn more about crypto and blockchain technology.

Glossary