Data SandboxTM
Your agents don't needbetter instructions.They need somewhere to fail.
- dbt run, SQL, backfills
- never writes to production
- bounded compute cost
real tables, read-onlyits own space to write4.20% of the bytes
[ production-scale data, without the risk or the bill ]
Somewhere to run the work, not just propose it.
Your agents have the skills and the MCPs. What they lack is a place to run dbt, SQL and backfills against real data. Set up in five minutes.
[ three jobs, one sandbox ]
Claude can already do this work.
Give it a place to run and it stops proposing changes and starts finishing them.
Improve the pipelines you already have
Claude reads your dbt project, rewrites the slow and fragile models, runs them, and compares the output against the original.
Runs
Build new pipelines, 6x faster
It writes the models, executes them on real data, finds its own broken joins and missing rows, and fixes them before you review.
Runs
Prove a hypothesis without building anything
Your team tests an idea against real production data first. Commission the pipeline only for the ideas that survive.
Runs
[ what's missing ]
Claude has the skills.
It doesn't have anywhere to use them.
Claude needs to run dbt run, fire real SQL, execute backfill scripts, and try a hundred variations cheaply. Your warehouse is the one place it can't safely do that.
[ The catch ]
It can't do any of it reliably on large data.
- A model that compiles can still be completely wrong
- So every idea needs a pipeline before it gets an answer
- And the best hypotheses never get tested
[ Why not just connect it ]
Giving Claude direct access to production tables is risky.
- Claude will try to query a table with billions of rows
- Claude can run your cloud bill up fast
- Claude can mess up your production data
[ how it stays safe ]
Real tables and real commands. Production stays untouched.
Read-only prod
It reads prod, it never writes it.
Claude reads your real tables through controlled, read-only access, then runs dbt, materializes models, and inspects the results in its own isolated scratch space. Keep your existing skills, MCPs, repositories, and warehouse; the sandbox is just the safe place underneath them.
production stays untouched
[ job 03 · in practice ]
Questions you can answer
before building a pipeline for them.
These need data that no clean table or semantic layer covers yet. Today each one is a pipeline request. Here, it's an afternoon.
- 01
What's the simplest change that would meaningfully increase user acquisition?
Growth - 02
Why did user growth slow down, and what should we test first?
Growth - 03
Which feature sets will most improve our churn prediction model?
ML - 04
How can we reduce fraud without blocking more good customers?
Risk - 05
Where can we cut data costs without changing a single output?
Cost - 06
Can this Snowflake workload run on Spark for less, with identical results?
Migration
Claude doesn't guess at these. It runs the numbers on real data, throws out what doesn't hold up, and comes back with an answer. Then you build the pipeline for the one hypothesis that held up.
[ read next ]
Stop building pipelines for everything.
Working on something like this? Chat with Founders →