// CONNECT YOUR AI · CURSOR

CONNECT CURSOR

Cursor reads MCP servers from a file called mcp.json. Add QuantGPT's address with a key in the header, restart Cursor, and its agent can run backtests and read-only warehouse queries while you work. Cursor's pricing page lists MCP on its paid plans, from Pro at $20 a month.

Updated September 2026 · Cursor facts from its maker's own docs, linked under Sources

// WHAT YOU GET

Once connected, Cursor can search QuantGPT's strategy library, run backtests on decades of US market history, grade a stock and query the research warehouse, in the middle of a conversation. The work runs on QuantGPT's side, and Cursor gets the numbers back, losing years included. The connection uses MCP (Model Context Protocol), the open standard AI apps use to reach outside tools.

// BEFORE YOU START

Cursor, on a plan with MCP. Cursor's pricing page lists MCPs, skills and hooks among the features of its Individual plans, from Pro at $20 a month. The free Hobby plan's list does not include them.
A QuantGPT account. The Free plan is enough to try it: 5 AI calls for the life of the account, no card. QuantGPT ($49 a month) allows 500 AI calls a day and QuantGPT Pro ($149 a month) allows 1,500. The Free plan can issue one key, QuantGPT three and QuantGPT Pro five. The plans.

// THE STEPS

Issue a key. Sign in at terminal.quantgpt.co and open your account page. Under Keys for your AI, type a label (for example cursor) and press Issue key. Copy the qg_sk_... key right away: it is not stored and will not be shown again.
Open your global mcp.json. Open ~/.cursor/mcp.json in your home folder (on Windows, C:\Users\YOU\.cursor\mcp.json), or create it. The global file makes QuantGPT available in every project.

Add QuantGPT. Paste this, with your key in place of qg_sk_YOUR_KEY. If the file already lists servers, add only the quantgpt entry inside the existing mcpServers object.

{
  "mcpServers": {
    "quantgpt": {
      "url": "https://terminal.quantgpt.co/mcp",
      "headers": {
        "Authorization": "Bearer qg_sk_YOUR_KEY"
      }
    }
  }
}

Save and restart. Save the file and restart Cursor.
Check it is on. Open Customize in the sidebar, then MCPs. QuantGPT should be listed, with its toggle on.
Ask the first prompt. Ask in the agent chat. Depending on your Run Mode, Cursor asks for your approval before an MCP tool call or runs it: in Cursor 3.6 and later, the default Auto-review mode runs allowlisted tools at once and sends the rest through a safety classifier. Click the arrow next to the tool's name to see the arguments.

For a shared project instead, put the entry in .cursor/mcp.json in the project folder, and never commit the key itself: write "Authorization": "Bearer ${env:QUANTGPT_KEY}" and set QUANTGPT_KEY in your environment. Cursor reads environment variables when it starts, so restart it after setting one.

// CHECK IT WORKED

Ask: Call QuantGPT's describe_capabilities tool and tell me what data it covers. A connected server answers with QuantGPT's own map: the data it holds, the factors, the benchmarks and the method. Cursor shows each call in the chat, with the arguments and the response you can expand.

// YOUR FIRST PROMPT

Backtest a 200-day moving average rule on SPY and show the losing years

Cursor's agent uses MCP tools on its own when they fit. You can also name one: use the quantgpt run_backtest tool.

Your AI picks the tool and the settings. It may test the rule on SPY itself, or on the S&P 500 index with the closest built-in rule (the 10-month average, or the 50-day and 200-day golden cross). Ask it to say which tool it called and with what settings, so you know exactly what was tested.

// WHAT COMES BACK

A backtest answers with the numbers behind it, not a summary. Your AI receives:

The rule as the engine read it, the dates covered and the number of periods.
For the rule and for its benchmark side by side (for a timing rule, the benchmark is buy and hold): growth per year, the worst fall from a peak, volatility, the Sharpe ratio, the share of winning periods, and a luck check (the probabilistic Sharpe ratio).
The same statistics split at a cut year (2015 unless you pick another), so the years after the cut read on their own.
A table by decade.
Every period's return, for the rule and for the benchmark. That is how your AI finds the losing years: it adds the returns up by calendar year.
For stock-picking backtests, a cost model: the measured turnover and the results net of 5, 10, 25 and 50 basis points a side.
A label from a fixed rule (confirmed, tradeoff or rejected) with the sentence that says what the label means.
Often, links to share-ready PNG charts, one of them year by year. Those links are public: anyone who has one can open it.
A slow run answers with a job id first; your AI collects the result with fetch_job.

Two limits travel with the numbers. Companies that later delisted are not in the market-wide backtest universe yet (a fix is in progress), so a backtest across the whole market leans toward survivors; a rule on one index, or on a list of tickers you choose, is not affected. And a backtest shows what happened in the past, not what will happen. QuantGPT's server instructions ask your AI to say so when it reports a market-wide result.

Cursor attaches any image a tool returns to the chat, so charts from QuantGPT's charts category show inline.

// THE TOOLS CURSOR SEES FIRST

On connect, QuantGPT lists 19 tools: 14 research tools, and 5 that switch on the rest by category (equities, portfolio, macro, alt_markets, events, charts and desk). The server also sends your AI instructions when it connects: what the data holds, the method, and how the tools fit together.

TOOL	WHAT IT DOES
`search, fetch`	Find and open QuantGPT's own records to cite: library strategies, study write-ups, forum threads, your saved studies and glossary terms, each with a link.
`describe_capabilities`	The full map: every factor, benchmark and timing rule, the library's categories, what each result label means, and the method. Your AI should call it first.
`guide, get_prompt_library`	A step-by-step course for new users, and ready-made prompts matched to the tools that answer them.
`search_library`	Search 175 documented strategies, 63 of them runnable.
`run_backtest, run_sweep`	Backtest criteria (up to 8 filters and a ranking, or a list of tickers you choose), or up to 100 variants in one call.
`challenge_finding`	The skeptic pass: eleven checks that try to break a result, strongest objection first.
`fetch_job`	Collect a slow run that answered with a job id.
`grade_stock, pre_trade_brief`	One stock's grades on six dimensions against the market, and a read that puts the strongest objection first. Never an order.
`query_warehouse, list_warehouse`	Read-only SQL over every research database, and the map of its tables.
`available_categories, available_tools, activate_category, deactivate_category, call_tool`	See the other categories, switch them on for your key, or run any tool by name.

Every tool can be called at any time. If Cursor does not see a tool your AI switched on, turn QuantGPT off and on again under Customize → MCPs, or ask the agent to run the tool by name through call_tool.

// TROUBLESHOOTING

QuantGPT is not listed under Customize → MCPs. Check that mcp.json is valid JSON (every brace closed, commas between entries) and in the right folder, then restart Cursor.
See what went wrong. Open the Output panel (Ctrl+Shift+U on Windows and Linux, Cmd+Shift+U on Mac) and select MCP Logs.
The logs show HTTP 401, or "invalid or missing bearer key". The header must read exactly Authorization: Bearer qg_sk_..., with one space after Bearer, and the key must still be active on your account page.
The ${env:QUANTGPT_KEY} value is empty. Cursor only sees variables that were set before it started. Set the variable where Cursor can read it, then restart Cursor.
Cursor runs calls without asking. That is your Run Mode. In Cursor 3.6 and later, it lives under Cursor Settings → Agents → Approvals & Execution. Every call counts against your plan's AI calls, so automatic runs can use the day's count quickly.
An error mentions the daily cap, HTTP 429 or HTTP 402. Your plan's AI calls are used up for now. The count is per account, across every key and connection; only tool calls count, not connecting or listing tools; and it resets at midnight UTC. HTTP 402 means the 3-day trial's 5 AI calls are used: start the plan from your account page. A message that starts with slow down means the limit of 120 requests a minute per account: wait a minute.
The answer says the engine is busy. QuantGPT runs backtests in a fair queue. Wait a few seconds and ask again.
To disconnect. Delete the quantgpt entry from mcp.json, or turn it off under Customize → MCPs, and revoke the key on your QuantGPT account page.

// LIMITS AND BOUNDARIES

AI calls are counted per account, across all your keys and connections: 5 for the life of a Free account, 5 during a 3-day trial, 500 a day on QuantGPT and 1,500 a day on QuantGPT Pro. Only tool calls count toward those; connecting and listing tools do not. Daily counts reset at midnight UTC. Separately, each account can make at most 120 requests a minute.

No QuantGPT tool can place, change or cancel an order. A few save things to your own QuantGPT account, such as a saved study or an alert. QuantGPT is research software: it reads market history and reports what happened, losing years included. Nothing here is investment advice.

// FAQ

Does Cursor support remote MCP servers like QuantGPT?

Yes. Cursor supports stdio, SSE and streamable HTTP servers, and QuantGPT's endpoint speaks streamable HTTP at https://terminal.quantgpt.co/mcp.

Do I need a paid Cursor plan?

Cursor's pricing page lists MCPs, skills and hooks among the features of its Individual plans, from Pro at $20 a month, as of September 2026. The free Hobby plan's list does not include them.

Can Cursor connect to QuantGPT without a key?

Cursor supports OAuth sign-in for remote servers, and QuantGPT's endpoint offers one. This guide uses a key because it is one entry in mcp.json and you can revoke it on your account page at any time.

Is my key safe in mcp.json?

It sits in the file as plain text. Keep it in the global file in your home folder, or use ${env:QUANTGPT_KEY} in a project file. If a key leaks, revoke it on your account page: a revoked key stops working on its next call.

What does QuantGPT cost?

The Free plan gives 5 backtests and 5 AI calls for the life of the account, with no card. QuantGPT is $49 a month with 500 AI calls a day, and QuantGPT Pro is $149 a month with 1,500 a day. Both paid plans start with a 3-day trial that includes 5 backtests and 5 AI calls.

Can Cursor place trades through QuantGPT?

No. No QuantGPT tool can place, change or cancel an order. QuantGPT is research software: it reads market history and reports what happened, losing years included. What you do with a result is your decision, at your own broker.

What data does QuantGPT cover?

US stocks and indexes (the S&P 500 from 1928, single stocks thin before 1980), fundamentals joined by filing date (about 9% of rows still carry the quarter-end date; a fix is in progress), crypto with funding rates, commodities, FRED macro series, prediction markets, sports odds and weather, plus a library of 175 documented strategies, 63 of them runnable. Companies that later delisted are not in the market-wide backtest universe yet, so market-wide results lean toward survivors.

// SOURCES

Facts about Cursor come from its maker's own public pages, as of September 2026, and may have changed since:

Cursor: Model Context Protocol (MCP): where mcp.json lives, the remote server format with headers, ${env:NAME} values, transports, OAuth support, tool approval, MCP Logs and images
Cursor: MCP integrations: Customize → MCPs, save and restart, the approvals setting and troubleshooting
Cursor: Pricing: MCPs listed among the paid Individual plan features, from Pro

Facts about QuantGPT come from QuantGPT's own server and plans.

CURSOR IS A TRADEMARK OF ITS OWNER; NO AFFILIATION OR ENDORSEMENT IS IMPLIED. QUANTGPT IS A RESEARCH TOOL, NOT INVESTMENT ADVICE. BACKTESTS ARE NOT PROMISES. THE TERMS GOVERN.