Ask ChatGPT or Claude about your options positioning. GammaGrid now has an MCP server.

Where are SPY's walls today? What changed for my tickers since Friday? Anything unusual in NVDA's option volume? Those are questions, and since today you can ask them where you already ask questions — in ChatGPT, Claude or Claude Code — and get the answer from the chains GammaGrid has collected for your watchlist, with the time it was collected and a link to the screen. It is read-only, it needs no API key, and it takes two minutes to connect. Here is what it answers, how, and where it stops.

MCP — the Model Context Protocol — is the open standard for plugging tools into an AI assistant. A server describes what it can do in a short list; the assistant reads that list, decides which tool answers your question, calls it, and writes the reply out of what came back. ChatGPT, Claude.ai, Claude Desktop and Claude Code all speak it. What has been missing is anything on the other end that knows about options positioning — dealer gamma exposure, walls, max pain, open interest — rather than just prices.

So GammaGrid has one now. It is the same product, the same stored rows the screens read, exposed as ten tools an assistant can call. Nothing is computed on the fly and nothing is fetched from a market feed when you ask; the assistant reads what the collector already wrote, exactly as the dashboard does, which is why every answer can say when it was collected.

What you can ask

Ten tools, all read-only. You do not call them by name — you ask in your own words and the assistant picks. The middle column is what it will pick, so you know what to expect back.

Every answer carries three things beyond the numbers: the sentence the assistant can say back, the structured object it can compute with if you ask a follow-up, and the signature — the data source, the collection time in UTC, and "approximate GEX, dealer sign by heuristic" wherever GEX appears. Those are the same words that stand under every chart in the app; the assistant does not get a cleaner version of the truth than the screen does.

How to connect it

One step in the app, then one step in your assistant. Sign in at app.gammagrid.io, open "Ask in ChatGPT or Claude" at the top of the sidebar and press the button. That turns the feature on for your account and shows you the server address and the exact instructions for each assistant — they live there rather than here because they name your environment's address, and because they will change as tools are added. In outline:

ChatGPT. Settings → Security and login → turn on Developer mode, then Plugins → add an app for a remote MCP server: paste the address, choose OAuth, set registration to dynamic client registration and leave the client fields empty — ChatGPT registers itself. Web, and on a paid plan; developer mode is not offered on the free one.

Claude.ai, Claude Desktop or the mobile app. Customize → Connectors → Add custom connector, paste the address, leave the Client ID and secret empty. The free plan allows one custom connector.

Claude Code. One command, which the sidebar prints for you to copy — it is claude mcp add --transport http with a fixed callback port and the server address. Then /mcp inside Claude Code to sign in.

Whichever you use, the first question opens a browser tab: you sign in with the same Google or GitHub account as the app, and allow the connection once. The assistant keeps its own token from then on. There is no API key, and nothing to paste but the address.

What it will not do

It will not write. No tool adds a ticker, changes a setting, or does anything a screen would call a button. If you ask it to add NFLX to your list, it will tell you where to do that yourself.

It will not answer about tickers you do not follow. Every per-ticker tool answers for the symbols on your watchlist and refuses the rest with a sentence naming where to add them. A new account is seeded with the free list, so it has something to say from the first minute.

It will not hand your assistant a whole chain. SPY is thirteen thousand contracts and about a megabyte; passed to a model, that is your whole conversation's worth of tokens spent on one question. So every tool has a ceiling built into it — so many strikes around the money, so many days back, the top so many rows — and the answer says what it trimmed. Ask a narrower question rather than a bigger one.

It is not real-time, and it is not advice. GammaGrid collects on a schedule during the trading day; the answer names the moment. What comes back describes dealer positioning the way the screens do. What to do about it is not in the data, and the assistant has no tool that pretends otherwise.

There is a daily quota per account — generous for a person asking questions, deliberately not generous for a script in a loop. The sidebar shows today's count.

What it costs, and who it costs

On our side, a call is a read of a stored row: single-digit milliseconds and a few kilobytes. On yours, a question is typically three to eight calls and a few thousand tokens of your assistant's plan — that is the tool descriptions, which the assistant reads once per conversation, plus the answers. GammaGrid does not run a model and has no key of yours; the assistant is the one you already pay for, or the free one. The feature itself is free while GammaGrid is in beta, like the rest of it.

Why an MCP server and not "an API"

Because the question people actually have is "where are the walls", not "what is the JSON schema for walls". An API is a promise to developers, and a good one takes rate plans, keys, versioning and documentation that stays true. An MCP server is a promise to a person with a question: the tool descriptions are the documentation, the assistant is the client, and sign-in is the key. It also happens to be where the audience already is — the assistant is open in the next tab. A conventional API may come later, and if it does it will read the same rows through the same ceilings; nothing here is a preview of something else.