TSN’s Richard Lofthouse found something useful while building a trading bot: a Grok Bot does not just write analysis code, it can also run it. Each Bot has its own cloud computer with a terminal, so it can install Python’s standard statistics libraries, fit models and keep the results, without the user’s laptop being on.
We tested this on 16 October. On a Grok Bot’s computer, the statsmodels library installed in about 14 seconds and fitted a simple regression in about a thousandth of a second. There was no GPU, and we found no published limit on the computing power available.
What the documentation says
The Grok Bot overview says each Bot “works on a persistent cloud computer with a browser, filesystem, and terminal”, and that work carries on when your laptop is closed. According to the page on the computer and apps, all the Bots on one account use the same computer and share its files, browser sessions and command-line credentials.
The security FAQ says each user gets a dedicated Firecracker microVM, a lightweight virtual machine with its own kernel and memory. One user cannot reach another user’s computer. A feature called Auto-review can let shell commands run, ask for approval first or block them.
What the test showed
The test was run on 16 October at about 15:43 UK time. The machine had Python 3.13.5 on Linux.
The first attempt, installing statsmodels into the system’s own Python, was refused with an “externally managed environment” error. That is a standard Python safeguard (PEP 668), not a Grok Bot restriction. The fix is to create a virtual environment, a private folder of packages, and install there. With that, statsmodels 0.15.0 and numpy 2.5.4 installed in 14.3 seconds.
We then fitted an ordinary least squares regression to 200 synthetic data points generated from the formula y = 0.5 + 2x plus random noise. The model estimated an intercept of 0.509 and a slope of 1.959, close to the true values, with an R² of 0.92. The fit took about 0.001 seconds.
That machine showed 8 virtual CPUs, about 15 GiB of memory, a 125 GB disk and no GPU. Those figures were seen on one computer on one day and are not a guarantee. We found no published CPU, memory, GPU or runtime quota in the documentation.
The toolbox for a trading bot
The libraries most traders reach for all install with pip. pandas handles price data and rolling features. statsmodels covers regression, stationarity tests, ARIMA and Markov-switching models for spotting market regimes. scikit-learn provides machine-learning models and TimeSeriesSplit for walk-forward testing. arch fits GARCH volatility models. Prophet forecasts trend and seasonality, though it was designed for business data and is often weak on prices.
For people building on the open-source Freqtrade bot, FreqAI is its built-in machine-learning layer. It trains a model for each pair you trade on features you define, retrains periodically in live and dry runs, and emulates that retraining in backtests. Its documentation says it can use “any machine learning library/method available in Python”. It also warns that FreqAI cannot be combined with dynamic VolumePairlists, and that the example strategy “is not designed to be run in production”.
What about your own PC?
Lofthouse’s other point is that a Bot can also connect to your own computer, so a machine with a GPU could widen the toolbox. The documentation partly supports this. Grok Bot’s approvals page describes running commands on “the Mac or Windows computer in front of you” as a separate capability. You can set it to “Ask every time” (the default), “Always allow” or “Never allow”. However, the documentation does not mention GPUs. Using a local GPU would therefore depend on your own machine’s set-up, not on any documented Grok Bot feature.
What’s still unknown
- What compute limits apply. None are published, and the hardware we saw may vary.
- How heavy jobs count against plan usage. The overview says usage resets weekly, but we found no per-compute pricing.
- How well GPU work runs through a connected PC in practice.
What to watch
- Any published quotas or GPU options for Bot computers.
- Changes to the local-computer execution settings.
- New FreqAI releases and changes to its requirements.
What this means
- You can test statistical ideas on a Bot’s computer in minutes, without setting up a server.
- Files and keys on the computer are shared between all your Bots, so use exchange keys that can trade but not withdraw.
- A good model fit is not a trading edge. Nothing here is financial advice.
Sources
- Grok Bot overview, docs.x.ai
- Use the computer and apps, docs.x.ai
- Security FAQ, docs.x.ai
- Approvals, security and privacy, docs.x.ai
- FreqAI documentation, freqtrade.io
- statsmodels documentation
- pandas documentation
- scikit-learn documentation
- arch documentation
- Prophet documentation
Related stories
- Quick Guide: Run Statistical Models for a Trading Bot on a Grok Bot (companion guide)
- Quick Guide: Tame a Flood of Newsletters With a Grok Bot Daily Report
- Musk Backs a Grok Bot a User Built to Flag Scam Emails
- SpaceXSI Bundles Grok, Grok Bot, Cursor and X Premium Into One “X Subscription”
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
