Connect Cursor to Market Data

Cursor speaks MCP, which means it can query real market data directly instead of you pasting CSV into the chat. Below is the exact configuration, verified on 17 August 2026 — plus an honest note about what "live" does and does not mean here, because the phrase promises more than this or any similar server delivers.

What is actually live, and what is not

Worth settling before you configure anything, because the answer changes whether this is useful to you.

So: this is a historical archive with a live bridge to your own chart, not a market data feed. If you need streaming prices to trade against, you need a broker API, and no free MCP server will substitute for one. If you want an agent that can reason over years of real price history and interact with the chart you are looking at, this is exactly that.

Install

A sparse clone, because the full repository carries roughly 543 MB of candle data you do not need — the server fetches data over HTTPS and caches it locally instead.

git clone --depth 1 --filter=blob:none --sparse https://github.com/varsansri/fxabsolute
cd fxabsolute && git sparse-checkout set mcp
cd mcp && npm install

Verified on 17 August 2026: that clone is 806 KB, and the full tool smoke suite passes against live data in 13.4 seconds.

Note the absolute path of the mcp directory — you need it in the next step. pwd will print it.

The Cursor configuration

Cursor reads MCP servers from a JSON file. There are two scopes, and picking the right one saves confusion later:

The entry itself is the same either way:

{
  "mcpServers": {
    "fxabsolute": {
      "command": "node",
      "args": ["/absolute/path/to/fxabsolute/mcp/src/index.js"]
    }
  }
}

If the file already exists, add fxabsolute alongside your other servers inside the existing mcpServers object rather than replacing it.

Verify it works

Restart Cursor, then check the MCP settings pane — the server should appear with its tools listed. If it does not, the two usual causes are a relative path and a missing npm install.

The fastest functional test is to ask something the model could not possibly answer from training data, so a plausible-sounding hallucination is easy to spot:

Using the fxabsolute tools, which UTC hour has the widest
average range on EURUSD, and how many bars is that based on?

A correct answer names 14:00 UTC at about 2.83 pips, from roughly 80,000 bars in that bucket, against a quietest hour of 21:00 at 0.73 pips. If the model answers without citing a sample size, or gives round numbers, it is guessing rather than calling the tool.

You can also confirm the server independently, outside Cursor:

cd mcp && npm run smoke

The tools you now have

The design point worth understanding: aggregation happens server-side. Asking which hour is most volatile touches 1.9 million bars — as pasted CSV that is millions of tokens, and the model would have to do the arithmetic itself. Over MCP the server computes it and returns 24 rows. The token comparison across all four delivery methods is here.

What to actually ask it

Cursor's strength here is that it writes and runs code, so the useful pattern is let the agent fetch and compute, and review its reasoning rather than trusting a narrative answer.

• "Pull EURUSD H1 for 2024 and test whether Monday gaps fill within 24 hours. Show the code and the counts."
• "Compare average range in the London and New York sessions on XAUUSD. State the sample size for each."
• "How often did price touch the prior day's high and then close below it? Define 'touch' explicitly."

Two habits that separate useful output from confident nonsense: always ask for the sample size, and always ask it to state its definitions. Most wrong answers in this domain come from an unstated definition of "touch" or "gap" rather than from bad arithmetic.

And the part an agent cannot do for you: recognising a setup in real time is a trained human skill. FXAbsolute replays the same archive bar by bar in a browser, free with no account, so the agent handles the statistics and you build the screen time.

The Cursor config, in full

Cursor reads MCP servers from a JSON config. Since the server is published on npm, the entry is three lines and needs no local checkout:

{
  "mcpServers": {
    "fxabsolute": { "command": "npx", "args": ["-y", "fxabsolute-mcp"] }
  }
}

Restart Cursor, and the ten fxa_* tools become available to the agent. The same block works unchanged in Claude Desktop and Windsurf — it is the standard stdio form.

One thing to know about the first call: the server fetches the instrument file it needs on demand and caches it under ~/.cache/fxabsolute-mcp. The first question about a given instrument therefore takes a few seconds and roughly 38 MB; every question after it is local and instant. Nothing is fetched for instruments you never ask about.

Frequently asked questions