MCP Trading Bot vs a Hosted Platform in 2026: What an AI Chat Plan Plus a Broker MCP Server Costs a Month, and What It Leaves Out

Nine US brokers and exchanges now publish an official MCP server. That means a $20-a-month Claude or ChatGPT plan can read your positions and, at most of them, place an order, which was not true a year ago. It has produced a question we hear every week: why pay for a trading platform when the chat app you already have can talk to your broker? This page answers it with November 2026 numbers. It prices the chat route, the API route and a hosted platform side by side, then lists what the chat route does not include. Some of that is money. Most of it is the thing nobody prices: who sends the next prompt at 9:31 on a Tuesday while you are in a meeting.

What the MCP route actually is

MCP, the Model Context Protocol, is the plug standard that lets a chat model call tools: fetch positions, get a quote, place an order. The October post on broker AI announcements explains the protocol itself. The trading version has two pieces. One is a chat client, Claude Desktop or ChatGPT. The other is a small server the broker publishes, which you either run on your own machine or connect to as a hosted service. You put your broker API key in the server's configuration, open the chat, and ask the model to do something. The model decides which tool to call and with what arguments, the server sends that to the broker, and the broker does it.

Who publishes a broker MCP server, as of November 4, 2026

This is the part that changed in 2026. We checked each broker's own developer pages and repositories today. The pattern in the official ones is worth reading twice: caps, dry runs, confirmation tokens and sandbox defaults. The brokers that built these expect the model to make mistakes and put the brakes inside the server.

• Alpaca: official, open source, on GitHub. Places stock, crypto and options orders including bracket and trailing stops. Paper trading is the default; you flip an environment variable to go live. Release 2.4.1 landed today. Keys go in the client's config file.
• Robinhood: official hosted server since May 27, 2026, with a dedicated agent account. Robinhood's own announcement says it "does not control, supervise, monitor, recommend, or audit these AI agents."
• Interactive Brokers: official hosted server since July 28, 2026, where the AI drafts an order and the client submits it; a funded "AI Partition" for autonomous trading followed on October 21, 2026.
• tastytrade: official, self-hosted only, documented here. 84 tools: 70 read, 2 write, 12 the docs label destructive. Every order needs a dry run first, which returns a confirmation token that expires in 60 seconds. There is a read-only switch and a default cap of $50,000 notional. The default environment is production.
• Webull: official local server, documented here. Defaults to the sandbox, caps orders at $10,000 notional or 1,000 shares, and tells you in bold never to share your App Key, App Secret or token with an AI model in chat.
• Public.com: official repository, places orders, no paper mode mentioned, "use at your own risk" in the readme.
• Coinbase: hosted "Coinbase for Agents" server since June 11, 2026, authorized through OAuth rather than a pasted key.
• Kraken: a local server that ships with the Kraken command-line tool. The trade service is marked "Dangerous" and each call asks you to approve it.
• Tradier: official, 22 tools, places orders, with a paper-trading flag.

Charles Schwab is the notable absence. There is no official Schwab server; the community ones carry a "not approved, endorsed or certified by Charles Schwab" notice. Schwab's CEO said on October 28, 2026 that the firm is blocking unauthorized agents from logging in on clients' behalf, that it will let clients' LLMs connect to Schwab's own servers under Schwab's control, and that "we do not allow them to move money." If your account is at Schwab, the chat route does not exist yet in any supported form.

The monthly bill, three routes

Route 1: a chat plan plus a broker MCP server

• Claude: Pro is $20 a month ($17 on an annual plan), Max $100, Ultra $200. Custom connectors work on every plan including free, which allows one connector.
• ChatGPT: Plus is $20 a month, Pro from $100. OpenAI's help pages say full write access to custom MCP servers on the web is available to Business and Enterprise customers; Pro can connect custom servers for read and fetch; developer mode is Plus and Pro. Check the current page before you count on placing orders from a Plus plan.
• The server: $0 at all nine brokers above. Optional: a $4 to $5 a month virtual machine if you want the server on a box that stays on, though the chat client still needs you, as the next section explains.

So the chat route costs $20 to $200 a month before the first order, and the plan's usage limit is shared with everything else you ask the model that day.

Route 2: the API, no chat plan

If you write your own loop against the model's API instead, you pay per token. An illustration, not a quote: one call every five minutes across a 6.5-hour session, 21 sessions a month, is 1,638 calls. At 3,000 tokens in and 500 out per call that is 4.9 million input tokens and 0.8 million output tokens a month. On Anthropic's published prices today, Sonnet 5.5 at $2 and $10 per million comes to about $18 a month; Opus 5.5 at $4 and $20 about $36; Fable 5.1 at $10 and $50 about $90. Add the $4 machine. A bigger prompt, more symbols or a news feed scales that linearly. What the token bill does not include is the code around it: the scheduler, the retries, the order reconciliation and, at Schwab, the token refresh every 30 minutes and the re-authorization every 7 days.

Route 3: a hosted platform

JorgAI's plans are $14.99 a month for Starter (3 auto-trades a day), $39 for Plus (8), $99 for Pro (15) and $199 for Elite (unlimited), on the pricing page. The AI is part of the plan. There is no Claude or ChatGPT subscription underneath it, no model API key to buy, no server to run and nothing to install; the FAQ answer on our site to "do I need an MCP server" starts with the word no. The system runs on JorgAI's servers through the market session for stocks and around the clock for crypto.

What the chat route leaves out

Someone has to send the prompt

A trading system is a loop: look, decide, act, repeat, every few minutes, all session, every session. A chat is a turn: you ask, it answers. The scheduling features that were supposed to close that gap come with their own fine print. Claude Desktop's local scheduled tasks state plainly that tasks "only run while the desktop app is running and your computer is awake. If your computer sleeps through a scheduled time, the run is skipped," and that closing the laptop lid still puts it to sleep. Claude's remote scheduled tasks run on paid plans without your machine, but a task that needs a local server only runs locally, and the self-hosted broker servers above are local. ChatGPT's scheduled tasks are capped at 5 on Plus and 15 on Pro and the help page lists Gmail, Slack and GitHub as the apps a task can use, not custom MCP connectors. Net: the chat route is a tool you pick up, not a system that runs.

The keys live on your laptop

Alpaca, Webull and tastytrade all have you put the broker key in the MCP client's config file or your shell environment. That is a plain-text file on the same machine you browse the web with. Webull's own documentation says never to paste the key into the chat. Alpaca wrote in September 2026 that a leaked key can submit orders, cancel open orders and liquidate positions, though not start a withdrawal. The safe-connection guide covers what permissions a key should have. On a hosted platform the credential sits encrypted at rest on the platform's servers and you revoke it at the broker in one click; at JorgAI it is encrypted before it is written and the broker page shows the revoke path.

The model is not a rules engine

The difference that matters most is also the least visible. A rules engine checks the same conditions the same way on every pass: is the daily spend under the cap, is the position count under the limit, is the stop in place. A chat model answers a question each time it is asked, and its tool calls are best effort by design. Anthropic's tool-use documentation says that when a required parameter is missing the model "might also infer a reasonable value" and that "this behavior is not guaranteed." OpenAI added a strict mode specifically so calls "reliably adhere to the function schema, instead of being best effort." This is why tastytrade built the 60-second confirmation token and Webull built the caps: the people who wrote those servers do not trust the caller to be deterministic, and they are right not to.

The vendors say what the product is for

Anthropic's Usage Policy effective November 12, 2026 defines a personalized recommendation "to buy, sell, hold, or allocate among specific investment products" as a high-risk use and requires that "a qualified person must meaningfully review" it. OpenAI's usage policies effective November 20, 2026 prohibit "automation of high-stakes actions in sensitive areas without human review," listing "financial activities and credit." Read plainly, a consumer chat plan is licensed for an AI that drafts and a human who decides. An unattended trading loop run through it is outside what either company says the product is for, which is a different thing from a platform whose product is the unattended loop.

Who watches the exit

Say the model bought at 10:05 and you closed the laptop at 10:30. At a broker that holds a bracket on its own servers, the stop is alive. At one that does not, or if the model never attached one, the stop exists only in the transcript. What happens to your stop-loss when the connection drops walks each broker. A hosted platform places the exit at the broker where the broker supports it and watches the rest from its own servers; that is the job it is being paid for.

The record

The chat route leaves you a conversation. A platform leaves you an order log with the rule that fired, the price at the signal and the fill, which is what you want on the day a trade goes wrong and the only thing a broker's support desk can act on.

Where the chat route is the better tool

None of the above makes the MCP route useless. It is excellent at the things a turn is good for.

• Asking questions about your own account in plain language: concentration, cost basis, what moved today.
• A one-off order you confirm yourself, with the dry run and the confirmation token doing exactly what they were built for.
• Learning a broker's API before you write anything against it.
• Paper trading on Alpaca's sandbox, which is the default in their server for a reason.
• Brokers that designed for it: Interactive Brokers' draft-then-submit flow and Robinhood's separate agent account are the chat route done on the broker's terms. Can an AI agent trade stocks for you covers where agents genuinely help.

Where JorgAI fits

JorgAI is route 3. You connect the brokerage account you already have at Schwab, Alpaca, Tradier, tastytrade or Webull, or the exchange account at Kraken or the Crypto.com Exchange, over the broker's own API. You set the rules once: how much it may spend in a day, how big a position may be, how many auto-trades a day, the stop and the trailing stop. The system scans, applies those rules and places orders at your broker on its own servers, so your laptop can be closed and your phone can be off. No AI subscription, no model key, no MCP server, nothing to install. The free plan shows the same system trading a simulated account before you connect anything, and the no-code guide is the longer walkthrough.

Quick answers

Do I need an MCP server or an AI subscription to use JorgAI?

No. The AI is included in the plan and runs on JorgAI's servers. There is nothing to install or keep running on your side.

Can ChatGPT or Claude trade my brokerage account?

At nine US brokers and exchanges, yes, through the official MCP servers listed above, within the caps and confirmations each one builds in. Not at Schwab, which has no official server and blocks unauthorized agents. Both AI vendors' usage policies require a human to review financial actions.

How much does an MCP trading bot cost per month?

The broker servers are free. The chat plan is $20 to $200 a month on Claude or ChatGPT. Running your own loop on the API instead is about $18 to $90 a month in tokens for a five-minute scan through the session, on today's prices, plus a $4 machine and your time.

Is the chat route cheaper than a hosted platform?

On the subscription line it can be: Claude Pro is $20 and JorgAI Starter is $14.99; Claude Max is $100 and JorgAI Pro is $99. The chat route does not include a scheduler that runs while your machine sleeps, a rules engine, server-side exit monitoring or an order log, and its vendor's policy does not cover unattended trading. Which of those you need decides it.

Which brokers block AI agents?

Schwab blocks unauthorized agents from logging in and plans to allow LLM connections only through its own servers with money movement excluded. Every other broker above publishes a server with its own limits.

Facts, prices and quotes are as of November 4, 2026 from each company's own pages and announcements, named in the text; prices change, so check the current page before you buy. The API-route figures are an illustration from stated assumptions, not a quote. Descriptions of JorgAI are of how it works, not of results. Education, not financial advice.

If you want the loop without the subscription stack, start a free trial: connect the broker you already use, set the rules once, and the system runs them on its own servers.