Questions before you start.

Takumis is in private beta. Register to request access. During the beta, Takumis builds and backtests strategies; running one on a trading account comes later.

No, and you never see any. You describe the strategy in ordinary language, and Takumis shows you the rules it built in the same ordinary language. You change them by asking, not by editing anything.

Three kinds of thing. You choose one when you start a project, and you can start a new project whenever you want, so the choice never locks you in.

• A strategy: rules that trade the same way over and over. This is the normal choice, and the right one if you are unsure.
• A workflow: a plan for a single trade around one upcoming event, such as a central bank meeting. It is a one-off plan that does not keep running.
• Mega Mode: you set a profit target and a budget, and the AI keeps trying on its own until it reaches the target or the budget runs out.

One project holds one strategy. To try a second idea, make a second project. The two stay completely separate.

Enough that another person could follow your rules without guessing. In practice that is four things:

• Which instrument, and which timeframe.
• What makes you enter a trade.
• What makes you exit, both when you are right and when you are wrong.
• How much you risk on each trade.

If something is missing, the AI asks instead of inventing it, because a guess buried inside the code is far worse than a question.

You do not need exact numbers up front. “Around 20 pips” is a fine starting point, and you can tune it after the first backtest.

Yes. There is a microphone button next to the message box. Press it, speak, then press stop. If the message box is empty, what you said is sent straight away. If you have already typed something, your words are added to it, so you can edit them before sending.

The first time, your browser asks permission to use the microphone. If you decline, the button will not work until you allow it in your browser's site settings.

Speaking is often the easier way to describe a strategy, because a strategy is a set of conditions and those are awkward to type but natural to say out loud.

Takumis says so, in plain language, and suggests the nearest thing that can actually be tested. And if you hold back part of the history, Takumis can test a change on data it was never tuned on.

58 instruments, all with the same seven timeframes and the same depth of history:

• 7 major pairs: EURUSD, USDJPY, GBPUSD, USDCHF, AUDUSD, NZDUSD, USDCAD.
• 21 crosses: EURJPY, GBPJPY, EURGBP, AUDJPY, CHFJPY and the rest.
• 28 exotic pairs: including USDMXN, USDZAR, USDSGD, EURPLN, EURCZK, EURHUF, USDNOK, USDSEK.
• 2 metals: gold (XAUUSD) and silver (XAGUSD).

No stocks, no stock indices, no crypto. We are working on adding more instruments.

To 2 May 2016. That is the earliest date a backtest can start, for every instrument and every timeframe.

That gives you ten years of data, which covers several very different market conditions, including:

• the 2016 Brexit vote
• the calm, steadily trending markets of 2017
• the trade war between the US and China in 2018 and 2019
• the COVID crash of March 2020
• Russia's invasion of Ukraine in February 2022, when oil rose above $100 a barrel and gold jumped
• the sharp interest rate rises of 2022 and 2023
• the broad US tariffs announced in April 2025, when the dollar fell sharply

The variety matters, because a strategy tested only on a quiet, trending period looks better than it really is.

It can run real calculations over real price data and explain the result in plain language. In practice that covers:

• How an instrument has behaved over a period: trend, range, volatility, typical daily movement.
• Where an indicator would have given a signal, and what price did afterwards.
• How two instruments move in relation to each other.
• How price behaved around economic releases.
• Statistics from your own backtests: which hours, weekdays or months worked and which did not.

It does not forecast prices. It describes what has already happened, which is far more reliable than a prediction.

Yes, although most people assume the opposite.

When you ask an analysis question, the AI does not answer from memory. It writes a small program, and that program runs against real historical price data, the same data the backtests use. The number you get back was calculated at that moment, from real prices.

A general-purpose chatbot asked the same question recalls something plausible. Here, every figure in an answer is calculated from the price data.

Yes, and many people use it mainly that way. You can ask questions in any project without ever asking for a strategy to be built.

Your project has an instrument and a timeframe set at the top of the screen, and a strategy is built and tested on that pair. If you ask about a different pair in conversation, the AI answers about that pair but does not change your project's setting.

This stops a casual question from switching your strategy to the wrong instrument without you noticing. To actually change what your strategy trades, change the setting at the top.

51 economic series across the eight currencies the platform covers: USD, EUR, JPY, GBP, AUD, CAD, CHF and NZD.

They cover the kinds of data that move currencies: inflation (CPI, PPI), interest rate decisions, growth (GDP), employment, activity surveys (PMI), retail sales, housing and sentiment.

Each release carries the date, the published figure and the surprise, meaning how far the number was from what the market expected. It is usually the surprise, rather than the number itself, that moves price.

Data for Japan includes the Bank of Japan rate decision, national CPI (headline and core), the GDP price index, and producer prices.

The prices are real. The trading is simulated.

A backtest replays your strategy over past market data, bar by bar, and records every trade it would have taken. It answers one question: if this strategy had been running over this period, what would have happened?

It runs on the broker's real prices for every minute of the past ten years, with spread and commission applied to every trade, and swap to every position held overnight. Every backtest keeps its full trade list, with the entry and exit time, direction, price and profit or loss of each trade, so you can check any result yourself.

But it is a simulation on past prices. It tells you whether an idea is worth more of your time. It cannot promise future results, and we will never present it as real trading.

Yes, all three. Every backtest charges the instrument's own average spread and a commission on every trade, and swap on every position held overnight. The profit figure you see is after all of those costs.

Each instrument is charged its own measured average spread. A single flat number for everything would badly misrepresent the wider pairs.

Swap is the overnight financing charged or paid each day a position is held, at 17:00 New York time, and tripled on Wednesdays. It is negligible for day trading and material for anything holding positions for weeks.

One cost is not included: slippage, the difference between the price you ask for and the price you get. The backtest assumes none. Live trading adds roughly 0.2 to 1 pip on majors in normal conditions, and far more around news.

We can't tell you that, and you should be wary of anyone who does. Takumis is a workshop: it lets you build and test an idea quickly and honestly. Whether the idea is any good is what the backtest is for. And a backtest is evidence, not a promise.

No, and it never will. The AI is a tool for writing and revising the strategy. Once the code exists, the AI is out of the picture. The strategy follows fixed rules that behave the same way every time, and no AI is involved when it opens or closes a trade.

No, on both counts. Takumis is software for building and testing strategies. We do not give investment advice, do not assess your circumstances, do not take or pass on orders, and do not hold your money. During the beta, Takumis only builds and backtests strategies; nothing runs on a trading account.

No. Not per trade, not on volume, and not for sending you to a broker. Nothing in the product has a reason to push you towards trading more, because trading more pays us nothing.