Connect Claude or ChatGPT to Your Trading Journal

What an MCP Connection Is

MCP, the Model Context Protocol, is an open standard for letting an AI assistant use tools in another application. The application publishes a list of tools, each with a name, a description and the inputs it takes. When you ask the assistant something that needs one, it calls the tool, gets the result back, and answers from it.

SignalDeck runs an MCP server with 54 tools. Connect Claude or ChatGPT to it once, and from then on a question like "which of my setups is losing money?" is answered from your actual trades, at the moment you ask, using the same code that produces the numbers on your dashboard. The assistant does not get a copy of your database. It gets the answer to each tool call it makes, and only for your own account.

A Paste Versus a Connection

Pasting an export into a chatbot works, up to a point. These are the differences that matter once you rely on it:

The arithmetic row is the one that gets underrated. A language model asked to compute a profit factor from 300 pasted rows will usually produce a plausible number, and sometimes a wrong one. A connected assistant calls a tool that returns the profit factor the app already computed, so the figure in the chat matches the figure on your dashboard.

What the Assistant Can See

Every connection includes read access. With it, the assistant can look up:

• Trades. Your trade list, filtered by status, symbol, strategy, date range or instrument type; full detail on any trade, including its notes, post-mortem, tags and change history; and every open position.
• Statistics. The dashboard numbers for the last 7, 30 or 90 days, the last year or all time: win rate, profit factor, average R, drawdown, expectancy, SQN. Also per-strategy performance, and monthly and yearly calendar summaries.
• Review. The review queue of trades that still need a note, a post-mortem or a stop, and your journaling and P&L goals.
• Risk checks. Portfolio heat across your open positions, a compliance check that projects a new trade against an FTMO preset, a generic preset or your own rules, and a concentration check by symbol and strategy. These only calculate. Nothing is logged.
• Backtests and strategies. Your backtest history and full results, your live strategies, and how each promoted strategy is tracking against its backtest.
• Journal and market. Your daily journal entries and mood, your tags, brokers and funding accounts, and the daily market review.

Each tool runs the same code as the matching screen in the app. A connected assistant on the Free plan sees what a Free user sees in the app, so SQN and the Kelly figure come back empty, exactly as they do on the dashboard.

What It Can Change, and How It Asks

Write access is a separate permission. You grant it when you connect, on the consent screen for claude.ai and ChatGPT or with a checkbox when you create a token for Claude Code, and you can connect without it. With it, the assistant can do the journal's data entry:

Three details matter here. First, a trade the assistant closes is computed by the same code as a trade you close by hand, so its P&L, R-multiple and slippage are identical. Second, if you edit a field on a trade that came from a broker sync, the assistant is told which fields came from the broker and warns you that the next forced re-import will overwrite them. Third, the hourly limits exist so that an assistant that misunderstands an instruction cannot rewrite your whole journal in one go.

On asking first: every write tool carries an instruction telling the assistant to show you the exact change and get your confirmation before making it, and most assistant apps also prompt you before running a tool that writes. Those are two layers, and in practice you see the change before it happens. But they live in the assistant and its app, not on SignalDeck's server, which cannot force an app to ask. That is the reason the permissions and the hourly limits exist as hard limits underneath. If you want a connection that cannot change anything, connect it read-only.

What It Can Delete, Only If You Opt In

Deleting is a third permission, on top of write. On the claude.ai and ChatGPT consent screen it is a checkbox that starts unticked, and for a Claude Code token you tick it when you create the token. Without it, the assistant has no way to delete anything. With it, it can delete trades, notes, tags, alerts, screenshots and journal entries, and each delete is limited further:

• Deleting a trade requires the assistant to pass back the trade's symbol. If it does not match, nothing is deleted. This guards against the assistant acting on the wrong trade id.
• A deleted trade cannot be restored. The assistant is instructed to name exactly what it will delete and to get your explicit confirmation.
• A tag that is still on a trade cannot be deleted, the same rule as in the app.
• Deletes count against the same 60-per-hour budget as trade writes.

Most people do not need delete access. Leave it off unless you have a specific cleanup job in mind, and revoke the connection afterwards if you only needed it once.

What It Can Never Do

No permission unlocks any of the following, because the tools for them do not exist:

• Place, change or cancel an order at your broker. The assistant works on your journal, not your brokerage account. Logging a trade in SignalDeck records it; it does not execute it.
• Touch billing, account settings or broker connections.
• See or change another user's data. No tool takes a user id. Every tool acts only as the account that signed in, so there is nothing to point at someone else.
• Change positions managed by the automated scanner. Those are refused even with write and delete access.

Three Things to Ask It First

The exchanges below are illustrative: the account and its numbers are invented, but the tool names are the real ones the assistant calls, and the answers use the fields those tools return. Your assistant's wording will differ.

1. Find out which setup is paying for the others

This is the question a single net P&L hides, and the reason win rate on its own is misleading. The follow-up is where a connection earns its keep: the assistant can pull the Breakout trades and check them, rather than asking you to export again.

2. Clear the review queue

Journaling usually fails as data entry, not as analysis. A review queue you can clear by answering questions gets cleared. Adding a planned stop that you really did plan is fine. Moving a stop after the fact to make a loss look smaller is not, and the journal keeps a change history either way.

3. Check a trade before you take it, then log it

The checks are read-only. Only the final step writes anything, and it records a trade you have already placed, or are about to place, at your broker yourself. For how the stop turns into a 1R, see fixed-R position sizing. Prop-firm traders can ask for the same check against an FTMO preset or their own firm's rules.

The Honest Limits

• The numbers are the app's. The reading of them is the model's. A tool returns a correct profit factor, and the assistant can still draw a weak conclusion from nine trades. If an answer matters, ask it which tool it used and what came back.
• What the assistant reads goes to its provider. Tool results become part of your conversation with Claude or ChatGPT and are handled under that provider's data terms, as well as SignalDeck's.
• Asking first is not enforced by our server, as covered above. Read-only is the setting for anyone who wants a guarantee.
• Large requests are split. A single result is capped in size, so "show me every trade I've ever taken" comes back in pages, or with a request to narrow the filter.

Setting It Up

You need a SignalDeck account; the Free plan works. The server address is the same for every client:

https://app.sgnldk.com/api/mcp

• claude.ai or ChatGPT. In the assistant's settings, add a custom connector and paste the address. Sign in to SignalDeck when prompted, choose read, write and, if you want it, delete on the consent screen, and approve. Custom connectors may need a paid plan or developer mode, depending on the provider and your plan with them.
• Claude Code. In SignalDeck, go to Settings, then AI assistants (MCP), and create a token, ticking write or delete if you want them. Then run:

  claude mcp add --transport http \
    signaldeck https://app.sgnldk.com/api/mcp \
    --header "Authorization: Bearer sd_pat_..."

• Claude Desktop. Create a token the same way, and copy the ready-made config snippet shown next to it into Claude Desktop's config, then restart the app.

Whichever you use, ask the assistant to run whoami first. It returns the account and permissions the connection has, which confirms you connected the account you meant to. For a step-by-step walkthrough, there is a short video on connecting Claude to SignalDeck, and the AI assistant page has the setup for each client and the full tool list.

To disconnect, go to Settings, then AI assistants (MCP). Revoke a token, or disconnect an app under Connected apps, and access stops at once. Changing or resetting your password also cuts off every connection. Logging out of the web app does not. Sign-in connections use short-lived access tokens that renew in the background, and a token for Claude Code is shown once, when you create it.

Plans and Limits

The connection is on every plan, including Free, and SignalDeck does not charge for it. Your assistant runs on your own Claude or ChatGPT account. Your SignalDeck plan's limits apply through the assistant the same way they apply in the app:

• On Free, SQN and the Kelly figure come back empty, and a trade can hold one screenshot.
• Plan-gated features, such as price alerts and the number of tags, stay gated. The assistant gets the same upgrade message the app shows.
• Screenshots can be JPEG, PNG, GIF or WebP up to 5 MB. iPhone HEIC files are not accepted, so export them as JPEG first.
• Each account can make up to 120 requests a minute, plus the hourly write limits above.

SignalDeck is free during beta. When billing launches, the plans are Free, Pro at $30/mo and Elite at $50/mo; see plans.

What to Do With This

Connect read-only first. Ask the three questions above and check the answers against your dashboard. When you trust what you see, reconnect with write access and let the assistant clear your review queue. That is the part of journaling most traders skip, and it is what makes every later question answerable in R.