← Coil home
GUIDE

tastytrade MCP and tastytrade AI Trading: set up your AI agent with Claude and Coil

tastytrade gives your agent hands. Coil gives it judgment. What tastytrade’s MCP server does, why the assistant that places an order is also the one that confirms it, and how to run both servers in one client.

Guide · 11 min read · updated November 2026

The one-sentence map: tastytrade provides the account and the order access. Your AI client provides the agent. Coil provides the read that decides whether, when and what to trade — served to your agent over MCP and never connected to your brokerage account.

Coil is not affiliated with tastytrade. Coil is also independent of Anthropic, OpenAI and the other AI providers named here. Everything below about tastytrade comes from tastytrade’s own pages: tastytrade MCP and the MCP server page of its developer docs, which the first links to, as read on November 15, 2026. tastytrade updates these pages, so check them for the current version. Nothing here is investment advice.

If you came for the connection, it is five steps below: connect tastytrade’s MCP, then add Coil’s in the same client, which is free and needs no key. If you came to decide whether an agent that can place orders is enough on its own, start with why an agent with hands still needs judgment.

What tastytrade shipped

tastytrade’s MCP page says you can “connect Claude, Cursor, or another MCP client to a tastytrade account.” It runs on the same Open API that powers tastytrade’s own platforms, so it has the same data, accounts and rules. tastytrade lists six areas the assistant can work in:

  • Market data: live quotes, snapshots, metrics, dividends and earnings.
  • Instruments and option chains: equity, option, futures and crypto instruments, and a full option chain in one ask.
  • Accounts and balances: balances, buying power and net liquidating value across every account your credentials reach.
  • Positions and transactions: open positions and transaction history explained in plain language.
  • Orders: a built-in review order ticket. tastytrade says every order shows a 60-second preview, and from there you can submit, edit, replace or cancel it, single-leg or complex.
  • Watchlists and alerts: watchlists and quote alerts the assistant maintains for you.

Confirmation, in tastytrade’s own words

The top of the page says nothing moves without your credentials and confirmation, and that “the assistant can suggest, but you decide what executes.” Its FAQ is more specific about the limits, and that is the part worth reading twice:

  • “Confirmation isn’t a second pair of eyes. The same assistant that places the order also confirms it, not a person reviewing it first.”
  • Some actions skip confirmation entirely and happen right away: canceling an order, deleting a quote alert and changing a watchlist.
  • The connection is built for one person at a time. It is not designed for managing multiple accounts or users at once.

tastytrade’s developer docs describe the mechanism. The five tools that submit or change an order each take a single-use confirmation token, minted by that tool’s own dry run and valid for 60 seconds. The tools that cancel an order, delete a quote alert or change a watchlist have no token and no dry run. A token that the agent mints and then spends itself is a guard against a slip, such as a wrong quantity. It is not a person agreeing to the trade.

Where the server runs

The two pages describe the product differently. The MCP page says it is “a connection, not an app” and that you don’t install new trading software. The developer docs say the server is self-hosted and open source, that you run it yourself, and that there is no tastytrade-hosted endpoint to connect to. They give clone-and-build as the supported install and require Node.js 22 or newer. Use the docs for the install, since they are the ones with the steps.

Credentials, the sandbox and the read-only setting

The MCP page says the server authenticates with OAuth2 rather than your login, using a client ID, client secret and refresh token you generate in your account settings, and that access tokens are short-lived, good for 15 minutes each. The docs add three settings that matter for a first run:

  • The environment. With nothing set, the server connects to production, where every order moves real money. A setting selects the sandbox instead. The sandbox needs its own, separate OAuth credentials.
  • Read-only. A setting withholds and refuses every write and destructive tool.
  • An order ceiling. A setting caps an order’s buying-power impact; the docs give a default of $50,000.

tastytrade says the sandbox does not serve live market data and resets every 24 hours, so it is for testing the connection and the order flow, not quotes. Its docs also say it recommends testing with paper trading before connecting a live account.

tastytrade’s notice on the page is plain: it does not endorse or recommend any third-party AI system, it is not responsible for what that system outputs or trades, and “language models are non-deterministic and may take actions their operator did not intend.”

Why an agent with hands still needs judgment

tastytrade’s MCP answers one question well: can my agent act on my account, through the same Open API my platform uses? It does not try to answer the other one: should it act now, on this name, at this price? That is a market question, and a preview ticket does not answer it.

A preview checks the order, not the idea

tastytrade’s review ticket and sanity checks look at the order itself: its size against account limits, its effect on buying power, whether the account can take it. They cannot tell whether the stock is at an entry or a chase. Coil scores every name on whether the price is inside its entry window or already extended. A name far above its entry is a chase, and the rules refuse to chase. A name that is falling is not bought at all.

Options chains are not a view

The assistant can pull a full option chain in one ask, which is a great deal to read. A chain tells an agent what is available, not whether the underlying’s trend is still intact. When the trend behind a holding is broken, the reason for owning the name is gone whatever the position reads, and exits come before entries. Coil’s hold read answers that question for each name, so an agent can act on a broken trend instead of waiting for a number.

Holding cash

An assistant that can always build an order has no way to say “do nothing today”. Coil’s market read does: it tells the agent whether the day is worth trading, and on a cash reading nothing new opens. Cash is a position, and an empty list on a given day is the discipline working.

None of this makes an order less your responsibility. A read can make an agent’s choices more consistent from one session to the next. It cannot see an order after it is placed and it cannot cap a loss. The settings you choose at tastytrade, and the amount you fund, do that. Use both.

Step by step: tastytrade’s MCP and Coil’s MCP in the same client

Step 1 — Decide where you start

The server needs a tastytrade account. tastytrade says opening one is free and that you do not need to fund it just to set up the connection. Decide before you install whether the first sessions run in the sandbox, in production with the read-only setting on, or both. In the sandbox no real money moves, and with the read-only setting on no order can be placed at all, so you can learn what the agent does first.

Step 2 — Install and connect tastytrade’s MCP from its docs

Use tastytrade’s developer docs for the install, because tastytrade keeps them current. They cover the clone-and-build steps, how to create the OAuth client and refresh token in your account, the environment variables and a client configuration for Claude Desktop and Claude Code. Two cautions from the docs: sandbox and production credentials are separate and do not work on each other’s side, and a shareable project-level config file is the wrong place for a client secret or refresh token. Never paste your tastytrade password into the agent or into a third-party site. Then verify with a read about your own account before anything else: “What are my current positions?” or “What orders did I place today?”

Step 3 — Add Coil’s MCP in the same client

Coil’s server is a second MCP connection, free, with no key, wallet or account. The address is:

https://mcp.coil.trade/mcp

In Claude Code, one command adds it:

claude mcp add --transport http coil https://mcp.coil.trade/mcp

In a desktop or web client, add it as a remote MCP server by URL in the connector or settings screen; the connect page has the one-click installs and a panel for each client. Then ask the agent: “What is the stock market doing, and is it worth trading?” A clean answer from Coil’s server confirms the connection. Coil’s server has no brokerage tools, so nothing on it can reach your brokerage account.

Step 4 — Give the agent a read order

Both servers are now in one client, and the agent will use whichever it is pointed at. Point it. The prompts in the next section fix the order: Coil’s read first, tastytrade’s tools second. Run it read-only or in the sandbox for a few weeks of sessions, asking the agent what it would have done and why, before you allow an order.

Step 5 — Allow orders, and be the second pair of eyes yourself

When the behavior is boring, you can allow the write tools on one account. tastytrade’s own words are that the assistant that places an order also confirms it, so no one else is checking. Write the approval into your instructions (“show me the preview and wait for my reply”), then read every preview against what Coil said before you reply. That step is an instruction, not a lock, so the settings and the amount you fund still do the real limiting. Fund the account only with cash you could lose.

Example prompts for an agent using Coil’s reads

Start with prompts you run yourself, with every order waiting on you. Each fixes the order of the reads, so the agent behaves the same way every session.

Before any trade idea: “Ask Coil whether the market is worth trading today. If Coil reads cash, build nothing new and tell me why.”

Looking for entries: “Ask Coil which stocks are at an entry. Compare them with my tastytrade watchlists. Skip any name Coil calls extended or falling, and tell me why for each one you skip.”

Checking what I hold: “For each position I hold at tastytrade, ask Coil whether the trend is still intact. Show me a sell preview for any broken trend and wait for my reply. Exits come before entries.”

Before buying one name: “Ask Coil for a read on this ticker. If its price is above its entry window, tell me it is a chase and do not build the order.”

When you are ready to write standing instructions for the agent, write them as rules it can check one by one, in this order:

  1. Read Coil’s market read first. If it says cash, make no new orders.
  2. Buy only a name Coil places inside its entry window. Never buy a name Coil calls extended or falling.
  3. Before every hold, ask Coil whether the trend behind the position is intact. A broken trend means sell it.
  4. Show me the order preview and wait for my reply before you submit anything.
  5. If Coil’s server does not answer, make no new orders.

Two things stay in your half of the instructions. Position size and the most you will allow on one order are yours to write, because Coil’s free reads publish scores and states, never position sizes. The Coil Live feed adds each name’s structural stop and target and the rulebook that sizes a position on that stop, refreshed about every 5 minutes during market hours; on Live, the agent’s opening prompt shrinks to one line, “call get_live_setup and follow it.”

FAQ

What is the tastytrade MCP server?

tastytrade describes it as a way to connect Claude, Cursor or another MCP client to a tastytrade account to ask about balances, check the market and review positions. It runs on the same Open API that powers tastytrade’s platforms. tastytrade’s developer docs add that the server is self-hosted and open source, and that there is no tastytrade-hosted endpoint to connect to.

Can Claude trade my tastytrade account?

tastytrade names Claude Desktop and Claude Code as the most common clients, and says Cursor, Zed and Continue connect the same way. With the write tools available, the agent can build an order, preview it and submit, edit, replace or cancel it. tastytrade says: “The same assistant that places the order also confirms it, not a person reviewing it first.” The agent has the access and no view of its own on whether a trade is worth making. That view is the part Coil supplies.

Does tastytrade’s MCP confirm each order with me?

Not with a person. tastytrade’s page shows a review order ticket with a 60-second preview, and its docs say the order-submitting tools need a single-use confirmation token from their own dry run. But tastytrade says the same assistant that places the order also confirms it, and that cancelling an order, deleting a quote alert and changing a watchlist skip confirmation and happen right away. Any step where you approve first is an instruction you give the agent.

Can I try it without real money?

tastytrade says there is a sandbox for testing. It does not serve live market data and resets every 24 hours, so it suits testing the connection and the order flow rather than live quotes. Its docs also describe a read-only setting that withholds every write and destructive tool.

Does Coil connect to my tastytrade account?

No. Coil’s MCP server has no brokerage tools and never sees your account. Your agent reads Coil for the market read and places any order through tastytrade’s own connection. Coil never holds accounts or keys.

What does it cost to add Coil?

Coil’s free MCP server needs no key, wallet or account: a delayed sample of the board’s morning read. Coil Live, the full read your agent calls over the same kind of MCP connection, is $49/mo or $419/yr, with a 7-day free trial. Nothing is installed. tastytrade’s own terms and fees are on tastytrade’s site.

Is letting an AI agent trade my tastytrade account risky?

tastytrade says language models are non-deterministic and may take actions their operator did not intend, that it does not endorse any third-party AI system and that you are solely responsible for the trading decisions you make with one. Coil does not remove that risk and cannot see or stop an order. A read can make an agent’s decisions more consistent; the settings you choose and the amount you fund are what cap the damage. You can lose money, so fund the account only with cash you could lose.

Does it work with clients other than Claude?

tastytrade says any MCP-compatible client works and names Claude Desktop, Claude Code, Cursor, Zed and Continue. Coil’s MCP server is one URL that any client taking a remote MCP server can add, so the same two-server setup applies. The steps differ by client; each client’s own documentation covers its own version.

Your agent has hands. Give it a read.

Add Coil’s MCP server to the client that already talks to tastytrade. It is free, with no key or account: https://mcp.coil.trade/mcp. Coil Live adds the full board, each name’s structural stop and target and the served rulebook, refreshed about every 5 minutes during market hours ($49/mo or $419/yr, nothing installed). Coil is not affiliated with tastytrade.

Try Coil Live free for 7 days

Coil is software and impersonal research, not investment advice. It never sells a stock short. On down days it may buy an inverse ETF, as a hedge or as a same-day trade. It is not a managed account and not a signal service. Leveraged ETFs, where the engine uses them, can lose value rapidly, including total loss. Backtests are research, not returns; the live record is Coil's engine trading one real brokerage account, published win or lose, not a projection of yours. Past performance does not predict future results.