How to Connect Your Brokerage to an AI Assistant (ChatGPT or Claude) in 2026

Guides for each major broker, and two group guides for the case that is actually most common: accounts spread across more than one.

• Fund-heavy accounts. Fidelity, Vanguard or Schwab, where fund overlap and asset location are the interesting questions.
• Single-stock accounts. Robinhood, E*TRADE or Webull, where position sizing, cost basis and employer stock matter more.
• One broker at a time. The per-broker guides are listed at the foot of this page.

Every assistant below reaches the same brokerage data through the same connector; what differs is where you paste the URL and what each client can do with the result. Start with the one you already use.

• Claude. Set up the Walnut connector (Desktop, claude.ai and Claude Code), or read the neutral survey of every route.
• ChatGPT. Connecting a brokerage to ChatGPT, which needs Developer Mode on a paid plan.
• Perplexity, Gemini, Grok and Copilot. Any client supporting remote MCP servers uses the same URL. See the connector reference for the current client list.
• Codex and Cursor. Add the server URL under MCP servers, then sign in.

For the whole capability, including what the assistant can see and the confirmation step before any order, see how Walnut works.

There are exactly three, ordered here from least to most technical. The first covers the large majority of people; the second suits power users who already work inside Claude or ChatGPT; the third is for developers.

• 1. Purpose-built apps. A hosted app links your existing broker and adds an AI layer you talk to, with no code. Examples: Walnut, PortfolioPilot, Magnifi, Mezzi, Origin. Best for most people.
• 2. MCP connectors. A Model Context Protocol connector exposes your brokerage to an assistant like Claude or ChatGPT directly, so the assistant itself can read your account. Examples: Walnut (many brokers at once), and the official connectors from Interactive Brokers, Robinhood, Webull, Public, tastytrade, Moomoo and Alpaca (each for its own accounts). Best if you already live inside Claude Desktop or ChatGPT.
• 3. The developer route. You wire a broker API (SnapTrade, Alpaca, Plaid) into your own agent for full control. Best if you write code.

Two things that look like routes but are not: a robo-advisor (Betterment, Wealthfront) holds your money itself and does not expose it to an outside AI assistant, and Claude used alone, with no connector, cannot see your holdings at all. ChatGPT's Finances feature is a partial exception for US users: it reads linked accounts but cannot place trades.

Whichever route you take, the core sequence is the same. For most people it takes a few minutes through a purpose-built app.

1. Pick a broker and an AI tool that support the connection. Confirm your brokerage is on the tool's supported list for reading your account, and that the tool works with the assistant you want (a built-in AI, or Claude or ChatGPT through an MCP connector).
2. Authorize a read-only connection. The tool sends you to a regulated aggregator where you log in at your broker, not inside the AI tool, and grant read-only access, so your credentials stay with the broker and nothing can move money without your approval.
3. Confirm your holdings appear. Once linked, check that your real positions, balances, and allocation show up correctly, so the assistant is reasoning over your actual account rather than a generic example.
4. Start asking questions. Ask about your concentration, how a holding has done against the S&P 500, or what has changed, and get answers grounded in what you actually own.

Whatever route you take, the same handful of things decide whether a connection is both useful and safe. Check these first.

• Credential safety. The connection should use a regulated aggregator so your broker login is entered at the broker and never stored by the AI tool. Be wary of anything asking for your broker password directly.
• Read-only by default. The assistant should analyze your account without the ability to move money unless you explicitly enable and approve trades.
• Which assistant it works with. Some tools have a built-in AI; others let you use Claude or ChatGPT through an MCP connector. Confirm the one you want is supported.
• Broker coverage. Whether your specific broker is supported for read, and separately for trade. The two lists differ.
• Setup effort. A hosted app is a few clicks; the MCP and developer routes need more setup and, for the developer route, ongoing maintenance.
• Honest disclosures. The tool should be informational, not claim to be registered investment advice.

This is the route most people want. A hosted app links your existing brokerage, usually through a regulated aggregator, and adds an AI layer you talk to, with no code and no server to run. You keep your broker; the app reads it and, in some cases, places trades you approve. Below are six worth knowing, each on the same fields so you can scan across them.

How to connect a broker through a purpose-built app, step by step

The flow is nearly identical across these apps, because they lean on the same aggregators. Using Walnut as the worked example:

1. Create an account in the app and open the brokerage-connection screen.
2. Pick your broker from the list (most major US brokers are supported).
3. Log in at your broker. The aggregator opens your broker's own authentication flow, so your password is entered at the broker and never seen by the app.
4. Authorize read access. The connection is read-only by default, which is enough for analysis.
5. Start chatting. Ask about your holdings, performance versus the S&P 500, or concentration, through Claude, ChatGPT, or the built-in assistant.
6. Enable trade approval only if you want it. Execution stays gated behind your explicit approval on every order.

If you already work inside Claude or ChatGPT, an MCP connector lets the assistant itself reach your account, rather than you visiting a separate app. MCP (Model Context Protocol) is the open standard, introduced by Anthropic in late 2024, that lets AI assistants connect to external data and tools through scoped, permissioned access. A brokerage MCP connector exposes your account so the assistant can read positions and, in some cases, place trades. A broker's own connector covers that one broker; Walnut is one connection across many. If all your money sits at one broker with its own connector, that connector is a fine choice.

How to set up an MCP connector, step by step

The exact steps depend on whether the connector is hosted (like Walnut's) or self-hosted (like Alpaca's open-source server), but the shape is the same:

1. Get the connector. For a hosted connector, generate a key in the app. For an open-source server, install and run it yourself.
2. Add it to your AI client. Register the connector in Claude Desktop (or your ChatGPT setup) so the assistant can call it.
3. Authenticate. Link the brokerage account the connector reads, through the aggregator or broker API it uses.
4. Confirm the access scope. Check whether it is read-only or trade-enabled, and whether trades require approval.
5. Use it in chat. Ask the assistant about your positions; it calls the connector to pull live data.

If you write code and want full control over how the AI reaches a broker, skip the hosted layer and wire a broker API into your own agent. It is the most flexible route and the most work, because you own the setup, security, and maintenance.

• SnapTrade API. Multi-broker read and trade aggregation you wire into your own agent, the same plumbing hosted apps use.
• Alpaca API. Programmatic brokerage access for fully custom automation, with paper trading to test against before going live.
• Plaid. Account aggregation for read access across many institutions, widely used for the read side.

Most people do not need this route, because a hosted app or MCP connector already wraps these same APIs. Reach for it when you are building a product or a personal agent with behavior the hosted tools do not offer.

Claude, not on its own: by default it has no link to your brokerage, so asking “how is my portfolio doing” gets a generic answer. ChatGPT is different for US users who link accounts through its Finances feature (via Plaid): it can show balances, holdings and daily changes. It still cannot make trades, cannot work toward target weights or baskets, may miss holdings at institutions that share only balances, and outside the US sees nothing. For anything beyond reading, you connect an account through one of the three routes above.

You can paste a list of holdings into the chat as a workaround, but it is a one-time snapshot: prices go stale immediately, you have to keep re-pasting, and long lists invite arithmetic mistakes. A live connection through a tool like Walnut keeps the data current and lets the assistant pull positions and prices itself, so the analysis stays accurate.

It helps to understand the plumbing, because the mechanism is what makes connecting a brokerage to an AI assistant safe. You never hand your broker password to the AI tool. Instead the connection runs through a regulated aggregator, in a few steps:

1. You pick your broker in the app or connector (Schwab, Fidelity, Robinhood, Public, and most other major US brokers show up here).
2. The aggregator opens your broker's own login page. You authenticate directly with Schwab, Fidelity, or whichever broker holds your money, exactly as if you were logging in yourself, so your credentials are entered at the broker and never seen or stored by the app.
3. The broker issues a scoped access token back to the aggregator, not your password. That token grants read access to positions, balances, and transactions, and nothing more unless you separately authorize trading.
4. The AI tool reads through that token. When you ask the assistant about your holdings, it calls the aggregator, which uses the token to pull live positions and prices. The AI sees data, not your login.
5. Revoking access is one click. Because the token lives with the aggregator, you can disconnect at your broker or in the app at any time, and the read access stops.

A trade-enabled connection layers a second, separately granted permission on top of the same token, and reputable tools still gate every order behind your explicit approval. This aggregator model, read-only by default with credentials kept at the broker, is exactly what to look for in any tool.

This is the distinction that matters most for security. A read-only connection lets the assistant see your holdings, performance, and fees and analyze them, without the ability to move money. A trade-enabled connection can also place orders. Neither should ever require handing your broker password to the AI tool. The safe pattern, used by Walnut and other reputable apps, routes the login through a regulated aggregator so credentials stay at the broker.

The Model Context Protocol, the open standard introduced by Anthropic in late 2024, was designed around exactly this: scoped, permissioned access to external data and tools. Reputable brokerage integrations (Walnut, and the brokers' own official connectors) keep your login at the broker and offer read-only defaults or limits, so the assistant gets context without getting the keys. The practical checklist: confirm the tool uses a regulated aggregator, confirm it defaults to read-only, and confirm any trade access is gated behind your explicit approval before you connect.

The list for reading is long. The list for trading is short. Through a regulated aggregator, most major US brokers connect for read access, including Fidelity, Charles Schwab, Vanguard, E*TRADE, Merrill Edge, Robinhood, Webull, SoFi and Public, which is enough for an assistant to analyze everything you actually hold.

Trade execution depends on the route. Through Walnut, trades you approve work at E*Trade, moomoo, Public, Questrade, Stake Australia, tastytrade, Trading212, Wealthsimple and Webull (US and Canada); Robinhood, Schwab, Fidelity, Vanguard and Interactive Brokers are read-only. Separately, several brokers now run official connectors for their own accounts: Robinhood's Agentic Trading MCP can trade a Robinhood account, and Interactive Brokers' lets the AI draft orders that you submit. Coverage changes at short notice and can quietly degrade at an individual broker.

The practical consequence: assume read-only unless your specific broker is confirmed tradeable in the tool you are using, today, and treat any checked-in list of “brokers you can trade” (including one on a page like this) as point-in-time. A broker's own connector is a different route from an aggregator: it goes deep on one broker, while an aggregator-based tool like Walnut spans many.

Connecting is a read permission. Letting an AI agent trade is a separate one, and the second is much harder to get. Most of the confusion about “agentic” investing comes from mixing the two, so it is worth separating them cleanly.

• Read access lets Claude or ChatGPT see your positions, allocation and concentration, so its answers are about your real holdings rather than generic examples. Nearly every large US broker supports this through a regulated aggregator.
• Agent execution lets the assistant submit an order. Through their own official connectors, Robinhood, Webull, Public and Alpaca allow it on their own accounts, and Robinhood added in-app Robinhood Agents on September 30, 2026. Interactive Brokers stops at a draft you submit. Schwab and Fidelity have no first-party connector we know of.

Which one you get is decided by the scope on the token when you authorize the connection, not by what you later ask the AI to do. A read-only connection is a hard limit: an assistant cannot talk its way past it, which is precisely why read-only is the sensible default. Where execution does exist, the careful products keep a human approval gate on every order rather than letting the agent act unsupervised, because a model can misread a ticker or size a position wrongly and will do so confidently. The broker-by-broker breakdown lives in the best agentic trading apps guide.

People search for this in two different vocabularies. Some ask how to connect a brokerage to an AI assistant; others ask which providers offer brokerage integration with AI assistants. They are the same question, and the honest answer is that the market splits into three kinds of provider, each with different coverage.

• Assistant providers (the integration is done for you). Walnut, PortfolioPilot, and Magnifi ship a finished integration: you authorize your brokerage account once and then ask questions in plain English. Coverage spans most major US brokers for reading, a narrower set for trading. This is the route with no setup work and no code.
• Aggregation providers (the plumbing others build on). SnapTrade and Plaid supply the account-linking layer itself. They are business-to-business infrastructure rather than something an individual investor signs up for directly, but they determine how broad any given app's coverage is.
• Broker-side integration (the broker exposes its own). Interactive Brokers, Robinhood, Webull, Public, tastytrade, Moomoo, TradeStation and Alpaca run official connectors of their own. Coverage here is exactly one broker: the one you are already with.

“Seamless” is worth defining before you judge a provider on it. A smooth integration means you authorize once, your login credentials never reach the AI tool, access is read-only until you explicitly grant more, and holdings refresh without you re-linking every week. Ask about coverage for your specific broker, and separately about whether that coverage is read-only or trade-enabled, because the two lists are never the same. For the provider-by-provider breakdown, see brokerage to AI integration providers compared.

There is no single best route, because it depends on how technical you are and what you want the AI to do. Match the route to yourself. Below, the field is ranked inside each use-case, with the stronger fit first; Walnut appears in its own category (the non-technical app route), not crowned across the board.

We compared the routes and tools on five things that decide whether a connection is both useful and safe: credential handling (does your broker login ever leave the broker), default access level (read-only versus trade-enabled), which assistant it works with, broker coverage for read and for trade, and setup effort. We did not rank by raw feature count, because a connector that can trade but stores your credentials is not an upgrade over a read-only one that does not. Features, broker support, and pricing change, so treat the specifics here as a starting point and verify on each provider's site.

To be upfront, since this is our site: Walnut is the AI investing assistant that talks to the broker you already have and places the trades you approve. It is one option on the first route and one on the second, not the only one and not crowned overall. A broker's own connector covers that one broker; Walnut is one connection across many. It connects each of your brokerages securely, then lets you analyze and manage it by talking through Claude, ChatGPT, or a built-in assistant. It is also a portfolio-aware MCP connector for people who prefer to stay inside their AI client: Claude Desktop, ChatGPT, or Cursor can read your real connected portfolio and, where the broker supports it, place trades you approve, using either Walnut's hosted connector or a small downloadable local MCP shim. It is read-only by default, every trade needs your approval, and you keep your existing broker. As a hosted MCP connector for Claude, ChatGPT, and Codex (and any MCP client), you connect it by adding its MCP server URL, https://walnutinvest.com/api/connector/mcp, as a custom connector, then signing in with a free Walnut account and approving read-only access; it never places trades. Full per-client setup lives on the Walnut MCP connector page. Walnut is not an investment adviser. If you mainly want a written second opinion, PortfolioPilot is worth a look; if you want tax-aware tracking across several accounts, Mezzi fits; if all your money is at one broker with its own connector, that connector is a fine choice; and if you are a developer wiring up your own agent, the Alpaca or SnapTrade APIs are the route.

From a connected account you can ask about a specific stock, an ETF you hold, or a theme you want exposure to, against your real portfolio. For the wider field, see the best AI portfolio analyzers roundup, the best MCP connectors for brokerages, or which brokers have an AI assistant.

Claude cannot see your brokerage by itself, and ChatGPT can only read linked US accounts without trading, so there are three ways to go further. If you are non-technical, a purpose-built app (Walnut, PortfolioPilot, Magnifi) connects your broker in a few clicks. If you live inside Claude or ChatGPT, an MCP connector gives the assistant direct access: Walnut across many brokers, or a broker's own connector for that broker only. If you build software, wire a broker API (SnapTrade, Alpaca, Plaid) into your own agent. Across all three, the safe choice keeps your login at your broker through a regulated aggregator and defaults to read-only, with any trade access gated behind your approval.