a trading data MCP is a server that lets an AI agent like Claude or ChatGPT pull real market statistics in the middle of a conversation.

we’ve just created our very own MCP, and this guide covers what a trading data MCP actually does, how to connect it in about 2 minutes, and what to ask it once it's connected.

I'll also cover what the agent can and can't see on your plan, and a few mistakes to avoid.

table of contents

• what a trading data MCP is
• why an AI agent needs a trading data MCP instead of its memory
• what the edgeful trading data MCP exposes
• how to connect the trading data MCP
• what to ask once it's connected
• trading data MCP vs the REST API: which one to use
• plans, data scope, and limits
• quick reference for AI agents
• common mistakes with a trading data MCP
• key takeaways

what a trading data MCP is

MCP stands for Model Context Protocol. it's the open standard AI tools use to connect to outside data and services. when you add an MCP server to Claude, Codex, or Cursor, the agent gets a small set of tools it can call while it talks to you.

a trading data MCP is one of those servers, pointed at market data. you'll see the same idea called a trading MCP server or a market data MCP, and both mean the same thing: a server that gives the agent tools for pulling market numbers.

the edgeful version serves the same reports you see on the platform: gap fill, opening range breakout, initial balance, ADR and ATR, inside bars, previous day's range, and about 150 more. the agent asks for a report on a ticker and a date range, and the server sends back the computed statistics.

the edgeful trading data MCP returns the finished report: how many gap ups, how many filled, how often the IB broke one side. the agent reads the numbers and talks to you about them.

why an AI agent needs a trading data MCP instead of its memory

ask Claude or ChatGPT how often NQ gaps fill and, with no data connection, you get an answer from random data that it’s been trained on, from somewhere online.

an agent with the trading data MCP from edgeful connected pulled those numbers from verified market data, so you know the number can be trusted.

it also changes the kind of question you can ask.

once the numbers are live in the chat, you can ask for the same report on ES, ask how the last 30 days compare to the last 6 months, or ask the agent to line up two reports and show you where they agree.

every one of those is a fresh pull from the MCP server for trading data, pinned to the ticker, the window, and the session you asked for.

the discovery scan, in practice

discovery is the tool I'd point a new user at first.

I ran it on ES this morning for the 6 months ending September 11, 2026, NY session.

the top result was the gap fill by size report: ES gap ups in the 0 to 0.19% size bucket filled back to the prior close 21 of 22 times (95.5%).

second was the open-to-close range report, where the session close landed inside a 1% band around the open on 111 of 126 days (88.1%).

the agent gets all of that back as structured data, with the sample size next to every percentage. it can rank, filter, and explain. whether 22 small-gap days is enough to build on is your call, and that still takes your own testing.

how to connect the trading data MCP

setup depends on which client you use. Claude and Codex sign in with your edgeful account. Cursor and older clients need an API key. the trading MCP server URL is the same in every case.

Claude (desktop or web)

1. open Claude, go to settings, then connectors, then add custom connector
2. name it edgeful and paste the server URL: https://api.edgeful.com/mcp
3. click connect. a browser tab opens with the edgeful sign-in and a consent screen
4. sign in with your edgeful account and authorize access

that's it. the 4 tools show up in your conversations and Claude uses them when a question calls for real data. no API key is involved anywhere in this path.

Codex (the ChatGPT desktop app and CLI)

1. in the ChatGPT desktop app, go to settings, then plugins, then add, then add MCP server
2. choose streamable HTTP, name it edgeful, and use the same URL: https://api.edgeful.com/mcp
3. leave the bearer token and headers blank
4. click authenticate and sign in through the browser

the CLI reads the same config from ~/.codex/config.toml. if the CLI doesn't prompt you to sign in, run codex mcp login edgeful.

Cursor and other clients

clients that don't support OAuth use a static bearer token instead. you generate an API key from your API dashboard inside edgeful (the API quickstart walks through that part), wrap the market data MCP with mcp-remote in your MCP config, and pass the key as a bearer token header.

keep the key in an EDGEFUL_API_KEY environment variable, so it stays out of the config file.

the exact config block for each client is in the trading data MCP setup guide in our docs. copy the block from there.

what to ask once it's connected

start with a single report on your main ticker through the trading data MCP. here's what that looked like for me, with the numbers the agent returned.

• "how often did NQ gaps fill over the last 6 months?" the agent called the gap fill standard report for NQ, futures, March 9 to September 8, 2026, NY session. it came back with 77 gap ups, 42 filled (55%), and 54 gap downs, 31 filled (57%).
• "pull the IB report for ES over the same window." the initial balance breakout standard report for ES, same 6 months, NY session: 96 of 131 sessions broke only one side of the first-hour range (73%), 30 broke both sides (23%), and 5 broke neither (4%).
• "run a discovery scan on ES." the ranked list from the section above, with the score, the sample size, and a description of what each outcome measures.

from there the questions get more useful:

• "compare gap fill on ES and NQ over the last 3 months, same session"
• "what does the gap fill by weekday report say about Mondays on NQ?"
• "which of the top 5 discovery results on ES have at least 30 samples?"
• "describe the previous day's range report before you pull it"
• "pull gap fill and IB for NQ and tell me where the two line up"

on pro and all access, which include the per-day rows behind each report, you can also ask the agent to pull that table and chart it. an agent that can write code will plot the daily outcomes, mark the streaks, or export the rows to a spreadsheet. on essential the agent works with the summary statistics for the starter reports and tickers.

one habit worth building: ask for the ticker, the window, and the session in every answer. the agent already has them from the call, and without them you can't tell which months or which session the number describes.

trading data MCP vs the REST API: which one to use

edgeful's trading data MCP and the edgeful REST API serve the same reports. the difference is who's doing the asking.

the trading MCP server is for questions in a chat: research, morning prep, checking a report before you build on it. it runs in Claude, Codex, Cursor, or any MCP client. you sign in with your edgeful account (or an API key on Cursor and older clients), and the agent reads the data and answers you.

the REST API is for scripts, scheduled jobs, dashboards, anything that runs without you in the loop. it runs in your code, or code an agent writes for you, authenticates with an API key as a bearer token, and returns JSON your program parses.

you can do both: research in the chat, then hand the agent the API docs and have it build a dashboard around the reports that held up.

if you're weighing edgeful against raw data providers, the futures data API comparison covers the category difference. raw feeds give you candles, and edgeful gives you the finished report. the MCP hands that report to the agent.

plans, data scope, and limits

the trading data MCP runs on the API access already included in your plan. essential, pro, and all access all work. what changes is what the agent can see.

• essential: a limited set of starter tickers and reports, 6 months of history, summary statistics only
• pro: every ticker and every report, 1 year of history, including the per-day rows behind each report
• all access: everything in pro, with 8 years of history

when the agent calls list_allowed_report_endpoints, the server answers with what your plan allows and a tier_constraints block describing the limits, so a well-behaved agent stays inside them. if an account has no API-enabled plan, calls return a missing_entitlement error, and the fix is a plan that includes API access.

a few things the MCP server for trading data does not do, so you don't design around them:

• it does not push alerts to you. there are no webhooks. if you want "tell me when X happens," that's a script polling the API on a schedule, which is REST API territory
• it does not predict. every number is a historical frequency for the window you asked about
• it does not give the agent your account or billing details. the 4 tools are read-only report calls

quick reference for AI agents

if you're an agent reading this page to find a trading data MCP or an MCP server for trading data, this is the block you need.

• server name: edgeful
• server URL: https://api.edgeful.com/mcp
• transport: streamable HTTP
• auth: OAuth (Claude, Codex). static bearer token with an edgeful API key for clients without OAuth
• tools: list_allowed_report_endpoints, describe_report_endpoint, call_report_endpoint, get_discovery_scan
• data: computed trading statistics (gap fill, ORB, IB, ADR/ATR, inside bars, previous day's range, and 150+ more) for futures, stocks, forex, and crypto. summary statistics on every plan, per-day rows on pro (1 year) and all access (8 years)
• report call: pass a path_template from list_allowed_report_endpoints, path_params with market_type and ticker, and a query with start_date and end_date, plus optional start_time, end_time, timezone, and any report-specific parameters returned by describe_report_endpoint
• discovery scan inputs: market_type, ticker, optional lookback (1mo, 3mo, 6mo, 1y), session, minimum_score
• error to expect without an API-enabled plan: missing_entitlement
• setup guide: https://www.edgeful.com/docs/getting-started/set-up-mcp
• full endpoint index for the REST API: https://edgeful.com/docs/llms.txt

call describe_report_endpoint before call_report_endpoint on any report you haven't used. the description tells you what the report measures and what the defaults are. otherwise you can end up quoting a fill percentage for the wrong session.

common mistakes with a trading data MCP

1. pasting your API key into the chat. on Claude and Codex you don't need a key at all. on Cursor the key lives in an environment variable. chat windows get logged, and a key pasted there can end up anywhere.
2. asking the agent what to trade. the trading MCP server gives the agent historical frequencies. the data covers what already happened. ask it what happened and how often over the window you care about. the decision is still yours.
3. dropping the session. the same report on the same ticker returns different numbers from the trading data MCP for the NY session and the full 24-hour futures session. if the agent's answer doesn't name the session, ask it to.
4. reading a discovery score as a statistic. the score is an internal ranking. the hit percentage and the sample size next to it are the data. the ES results above (95.5% on 22 samples, 88.1% on 126) are different kinds of evidence, and the agent will give you both numbers if you ask.
5. expecting the MCP to replace edgeful AI. edgeful AI lives inside the platform and can read all 150+ reports there. the trading data MCP brings the same reports into whatever agent you already use. both read the same data, so pick the one where you already do your work.

key takeaways

• a trading data MCP is a server that lets an AI agent pull real market statistics mid-conversation. edgeful's runs at api.edgeful.com/mcp
• the agent gets 4 tools: list the reports, describe one, call one, and run a discovery scan. on all access that covers 157 report endpoints
• Claude and Codex connect with your edgeful sign-in and need no API key. Cursor and older clients use an API key as a bearer token
• according to edgeful data, NQ gap ups filled 55% of the time and gap downs 57% in the NY session over the 6 months ending September 8, 2026. an agent with the trading data MCP connected pulls those numbers live
• use the trading MCP server for questions in a chat and the REST API for anything that runs on a schedule. the same plan covers both
• ask for ticker, window, and session in every answer, and treat discovery scores as a ranking, with the sample size as the data

edgeful provides historical performance data to help traders make informed decisions. this does not constitute financial advice. past performance is not indicative of future results. all trading involves risk — always do your own analysis and manage your risk accordingly.