← FXAbsolute For developers

Connect your coding agent to FXAbsolute

Your agent should not have to read prices off a screenshot. This MCP server hands it the numbers instead: five years of 1-minute candles across 15 instruments, and a live bridge to the chart you are backtesting right now.

Tool access needs a data contract

CHECK 29

Make the agent state what it read, calculated, and could not observe

Connecting a coding agent to historical candles reduces transcription work, but it does not make the model a market oracle. Begin each task with an instrument and data inventory, then require the agent to show query dates, timeframe aggregation, sample size, missing fields, exclusions, and the exact calculation behind every conclusion.

Protect replay integrity separately. A tool may technically access bars beyond the chart position even when the prompt asks it not to. Use bounded date parameters or a server-side replay boundary for any exercise where future leakage would invalidate the decision.

Data boundary

Confirm available symbols, fields, coverage, gaps, timezone, aggregation, and whether prices are midpoint, bid, or ask.

Tool boundary

List each tool the client can call, its read or write effect, response limits, and whether it can reach future data or the live chart.

Claim boundary

Separate measured output, implementation assumptions, and model-generated explanations in the final answer.

1. Call the inventory tool before requesting analysis.
2. Ask for the query manifest and denominator with every result.
3. Run a small hand-calculated case to verify aggregation and units.
4. Save prompts, tool outputs, versions, and errors for reproduction.

A useful trading agent is an auditable calculator with explicit limits, not a confident narrator with hidden inputs.

Reviewed 24 Aug 2026 Inspect the full MCP data and tool contract

Install

One command. Node 18 or newer.

claude mcp add fxabsolute -- npx -y fxabsolute-mcp

For any other MCP client, add this to its config:

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

Ask it something immediately

Once installed, the data tools work on their own — you do not need a chart open. Try asking your agent:

• “What does the first 5-minute candle of the New York session do on SPX500 over the last six months?”
• “Which hour of the day does EURUSD actually move? Group it by hour in New York time.”
• “Has 1.1000 held as a level on EURUSD this year, and what happened when price got there?”

Pair it with your live chart

1. Start a backtest on fxabsolute.com.
2. Switch the right-hand panel to the AI tab.
3. Tell your agent: connect to fxabsolute with code ABC123, using the six characters shown.
4. Type in the panel, or press the microphone and talk.

Every message carries a snapshot of the chart: instrument, timeframe, replay position, the current candle, your open trades, running performance, and every drawing as real geometry. A trendline arrives as its height in pips, the bars it spans and its slope per bar — not as an image the model has to squint at.

The snapshot also states where you are in the replay and tells the agent not to reveal bars you have not reached. An assistant that spoils the next candle would make backtesting pointless.

The tools

The data

• 15 instruments — EURUSD, GBPUSD, USDJPY, USDCAD, AUDUSD, AUDCAD, EURJPY, EURGBP, GBPJPY, XAUUSD, SPX500, NAS100, US30, BTCUSD, ETHUSD
• ~1.9 million 1-minute candles per forex instrument, 2021 to 2026, from Dukascopy tick data
• Fetched once per instrument (~37 MB) and cached on disk, so the first query takes seconds and the rest are instant

Two things stated plainly

There is no volume. The record is 20 bytes — a timestamp and four float32 OHLC values. Volume was never stored, so no tool reports it and none quietly substitutes a proxy. Bar range is offered instead, and every relevant response says so.

Timeframes are bucketed on wall-clock time, not by array index. Forex data has weekend and holiday gaps, so grouping every 60 one-minute bars eventually stops lining up with a real clock hour and an “H1 candle” can straddle a weekend. Here bars sit on true period boundaries, and a gap produces no bar rather than a fabricated one. Session windows resolve in exchange-local time with daylight saving handled to the second, so the 09:30 bar is the 09:30 bar in both January and July.

fxa_session_scan also reports how many weekdays in your window produced no data at all — worth reading, because a scan can otherwise return healthy-looking statistics built on a fifth less data than you asked for.

Questions