Architecture

Open source · Apache 2.0. This section documents the open-source core. The Data Workers platform is covered separately - see What the platform adds.

One agent, one domain

Each agent is a standalone MCP server scoped to a single domain. Your coding agent talks MCP; the Data Workers agent translates that into work against your data tooling. Run one or run all nine - they're independent processes.

A shared view of the estate

Agents don't message each other. They read and write a shared model of your estate - assets, lineage, quality history, past decisions - and coordinate through it. The catalog agent resolves analytics.dim_customers in Snowflake, default.dim_customers in Databricks, and dim_customers in dbt to one asset, so an answer from one agent is legible to the next.

The practical effect: the incident agent's root-cause analysis can lean on lineage the pipeline agent recorded, without either agent knowing the other exists.

Swappable underneath

Storage, messaging, search and model access sit behind interfaces. The open-source core ships local implementations and the sample estate, so it runs with no external services. Larger backings are selected by configuration.

Bring your own model

Agents call whichever model provider you configure, with your key. There's no Data Workers inference service in the path and no markup on your model spend.

Human in the loop

The open-source agents observe, analyse and recommend. Acting on a recommendation is a decision a person makes.