Claude Cowork Runs a Real Python Sandbox. I Put 37 Million Rows of Mutual Fund NAV Data Through It.
Published: August 2, 2026
Claude Cowork runs a Python sandbox of its own, and it can install packages it needs. And a decently powerful sandbox too. With access to tools like duckdb, statsmodel, scikit-learn you can run pretty decent analytics with it.
Python sandbox available in Claude in Excel and PowerPoint also....and Claude Opus 5 is brilliant whether you are building stuff on Claude Code or running analysis with Cowork.
A fair amount is there already there - pandas, numpy, scipy, duckdb, matplotlib and yfinance among them...and when I needed scikit-learn and statsmodels, it added both in under a minute....and you can add polars, pyarrow and many others.
For sandbox specs - I had Cowork run a quick benchmark. These are approximate - It reports two cores of an AMD Ryzen 7 7730U with about 4 GB of memory and 10 GB of disk, and other parameters which suggest a SSD territory. I'll say pretty good for a sandbox.
I work with MF Data very often. e.g I had it download the India Mutual Fund NAV data set from tigzig.com .. the data I publish, a 175 MB parquet holding 37 Million+ rows going back to 2006, along with the scheme master. Ran some MF Category analytics computing risk, return and other KPIs. Was running pretty fast as Claude was sharing final results less than a minute per iteration. (The section below has the AI agent instructions to pull the same data into your Claude / AI App.)
Just keep in mind that packages that need compiling can time out while installing, xgboost did, while the lighter ones go in without any trouble.
For guides on working with Claude for Analytics: Claude for Analytics. By and large same thing for other AI cowork and copilots apps.
Pull these datasets into your own agent
If you want your own agent to pull these India Mutual Fund datasets, just copy paste the prompt below. ..and by the way these are live datasets powering the analytics tool on MFPro tigzig.com/mfpro
Go to api.tigzig.com, find the mutual fund API, and download the full NAV master
and the scheme master, both in parquet.
The data goes back to April 2006 and keeps schemes that have since matured, merged or wound up, so you are not looking at survivors only.
A row is not a fund. It is one plan and option variant, so the same fund shows up several times, Direct and Regular, Growth and IDCW. Around 8,500 codes publish daily, which is roughly 1,950 funds the way AMFI counts them.
And scheme names move. SEBI has been recategorising, so names and categories shift over time. Match on scheme code or ISIN rather than on name.
Data dictionary: tigzig.com/mfpro -> Docs
Why DuckDB and Parquet
Most of my analysis and modeling work now happens with DuckDB / Parquet. It also powers the backend of my data intensive apps on tigzig.com ... with data sizes going up to 16 GB & 230 Million+ records. All live and public on the site.
As regards the MF NAV dataset - the same NAV data is 3.8 GB as a plain CSV and about 175 MB as parquet, roughly twenty two times smaller before you have run anything at all.
Parquet keeps each column together instead of each row, so a query touching three of the five columns reads only those three off the disk, and a column of dates or scheme codes compresses far better than a mixed row does, which is most of why 37 million rows go through a 4 GB sandbox without complaining.
DuckDB does the statistics too, they are built into the SQL, which you can run as direct functions without having to do any looping / pandas / statsmodel.
Want to know more about using DuckDB for analytics? Just paste this into your AI:
go to tigzig.com and pull out all posts and guides on using duckdb for analytics