Connecting ChatGPT or Claude to your brokerage account: which brokers enforce the guardrails

MCP, the Model Context Protocol, is how an AI assistant connects to an outside service. Point Claude or ChatGPT at a broker's MCP server, sign in, and the assistant can read balances, research positions and, depending on the broker, place trades. StockBrokers.com's guide, written by Jessica Inskip, opened funded accounts at each qualifying broker, connected them to Claude and ChatGPT where supported, and ranked them on what the broker itself allows or blocks.

That design choice is the right one for readers too. The guide's reasoning: you may switch AI models next month, so a broker earns credit only when it rejects a bad order itself.

Who qualified

The guide required a first-party server, a live funded account rather than paper trading, and general availability. Seven of the 14 brokers it covers met that bar; all seven launched their servers in 2026, and none charges for the connection. Charles Schwab and E*TRADE were excluded because their only MCP servers are community-built projects on the brokers' APIs. The guide adds that four brokers shipped or expanded their servers within six weeks of first publication, so documentation changes fast.

The guardrails, broker by broker

Three models of control

• The AI never touches execution. Interactive Brokers ranked first because the agent writes an instruction that you submit yourself from an IBKR platform. An AI that misreads you cannot send the order.
• The broker sets hard limits. Webull's caps, whitelist and read-only switch, and tastytrade's buying-power cap and dry-run token, sit outside the model. They hold even if the AI misbehaves.
• Containment by account. Robinhood isolates trading in a separate funded account. The reviewer's caveat matters: containment applies to trading, not visibility, since the agent can read balances, order history and watchlists across all your Robinhood accounts.

The one design to understand before you use it

Public offers two different things. Its hosted MCP server lets Claude, ChatGPT or Perplexity trade conversationally with preflight checks. Its Agentic Brokerage, by contrast, executes a strategy you confirmed once, from then on, without alerting you before each transaction, according to the guide. Public's own Agents FAQ describes the approved plan as running 'the same way every time' while noting the conversation used to build it is not deterministic.

Not a broker, but the same plumbing

Charting platforms are adding the same connection. A Kagels Trading review notes that TradingView offers an official MCP server so you can connect your own AI such as Claude, included from the Essential plan, with trial plans excluded (plans on our TradingView tool page). The reviewer's warning applies to every setup here: your own AI can still miscalculate or set a level in the wrong place; the difference is that you build the checks.

A safe order of operations

• Start read-only where the broker offers it, or with a broker where the AI cannot submit orders.
• Prefer limits the broker enforces over instructions written in your prompt.
• Isolate money. A separate, small account is the simplest cap that cannot be prompted away.
• Keep per-order confirmation on, and know which products drop it after a one-time approval.
• Log what the agent did. If the broker keeps an audit trail, read it weekly.