Intro to SpiceGrinder

SpiceGrinder: synthetic test data that actually looks like your data

If you work with health records, financial transactions, insurance claims, or anything else regulated, you know the problem: you need volume test data, but you can’t use production data to get it. PII, PHI, financial sensitivity — the same rules that protect real people make it hard to build, test, and demo against realistic data. Maybe your product is new, or you’re introducing a new feature and that production data doesn’t actually exist yet.

Most synthetic data tools solve this by giving you generic mock data — names, addresses, Lorem Ipsum with numbers attached. That’s fine until your data has real structure that you need to model with fidelity: claim frequency that follows a Poisson distribution, severity that’s log-normal, correlated fields, business objects with actual format and checksum rules. Generic mock data doesn’t capture any of that, and hand-rolling a data generator for every project doesn’t scale.

SpiceGrinder is a small, fast engine for generating synthetic data that mirrors your data’s actual shape — statistical distributions, correlated fields, and real-format business objects (SSNs, credit cards, phone numbers — with a toggle for guaranteed-safe reserved ranges when you need it), composed into a model that’s just as easy to read as it is to run. Everything runs inside your boundary — nothing ever calls home or requires you to connect to an external server. And it’s built to fit into whatever process you already have, not replace it: run it from the command line, drive it from the GUI, or hand it to an agent — however you already work is how you can use SpiceGrinder.

Let’s look at a simple example: claim frequency and severity for an insurance book

  <root node="ClaimObservation"/>
  <nodes>
    <Append name="ClaimObservation">
      <input name="Frequency"/>
      <input name="Severity"/>
    </Append>
    <Poisson name="Frequency" lambda="3.5"/>
    <Lognormal name="Severity" mu="8.0" sigma="1.2"/>
  </nodes>
</dataset>

Run it, and you get exactly the same output every time you use that seed — deterministic by design, so your tests stay reproducible. (If you want true randomness each time you run, simply omit or suppress the seed.) Using that seed of “42”, the model above will always yield the following results:

5,1020.778556837366
3,1693.1643701755693
8,7302.257200899249
...

Not AI, but agentic-friendly. SpiceGrinder doesn’t generate data by asking a language model to make something up — every value comes from a real, seeded, auditable statistical process. But it’s built for a world where a lot of engineering work now happens through AI agents: it ships with an MCP server and a set of purpose-built skills, so any MCP-aware agent can discover what components exist, help you build a model, validate it, analyze its performance, and generate data directly — no wrapper scripts, no scraping CLI output. If you’re already working with Claude Code, Cursor, or anything else agentic, SpiceGrinder is a tool your agent can just use.

The free tier is fully functional — every statistical distribution, no volume cap, real determinism — with no sales call required, no time limits. Just download and use as much as you like. The pro tier is for power users, and adds complex and custom data types, a library of business objects, and the ability to embed SpiceGrinder as a service in your own systems.

Try the beta →