Your data team’s AI intern, with read-only access and a paper trail

What it does

DataFoundry is an open-source AI workbench that turns natural-language questions into structured data tasks across dozens of sources. It wraps the agent in a governed runtime where business definitions, schema context, and read-only guardrails replace blind SQL generation. The output is not just an answer but a replayable trail of evidence—tables, charts, and audit logs—that teams can inspect, branch, and reuse.

The interesting bit

Instead of treating a database like a text file, DataFoundry forces the agent to align on business semantics—what “GMV” actually means in your schema—before it ever runs a query. The result is materialized as reusable workspace assets, not ephemeral chat bubbles, making the agent closer to an operating foundation for data tasks than a conversational toy.

Key highlights

• Supports 28 datasource types out of the box, from PostgreSQL to ClickHouse to MongoDB.
• Self-hosted with read-only defaults, credential isolation, field masking, and full audit replay.
• Multi-model support via any OpenAI-compatible provider (Qwen, DeepSeek, GPT, etc.).
• Branchable, concurrent analysis sessions with checkpoint-backed semantic trace DAGs.
• Built on Mastra for agent runtime and AG-UI for event streaming.

Caveats

• The project labels itself “early but usable” and ships without Docker or Compose builds.
• The automated deploy path is Ubuntu/Debian only; Windows and macOS are limited to manual npm builds.
• The semantic Data Link component is external and must be wired up separately if needed.

Verdict

Data teams that need governed, auditable AI analysis across complex schemas should look here; if you just want a quick SQL chatbot or a coding agent that writes patches, this is over-engineered.