Self-hosted text-to-SQL that makes you test the agent first

What it does

nao is a framework for building self-hosted analytics agents. Data teams use a CLI to assemble a file-system-like context—schemas, documentation, business rules, even external tools and MCPs—and then deploy a chat interface where non-technical users ask questions in plain English and get SQL-backed answers with charts. It turns your warehouse into a conversational BI layer, but only after you’ve built and tested the agent’s worldview.

The interesting bit

Most text-to-SQL tools treat the LLM prompt as the whole product; nao treats the agent like deployable software. It versions the context, tracks performance over time, and forces you to write unit tests in YAML before users can even open the chat UI. That’s the part that should make BI engineers nod.

Key highlights

• File-system-like context builder that mixes schemas, docs, and MCPs in the agent’s working memory.
• Built-in evaluation suite: define YAML test cases (question + expected SQL) and run them before deployment.
• Claims to be data-stack and LLM agnostic, though the README doesn’t enumerate specific connectors.
• Self-hosted by default with Docker support; you bring your own LLM keys and keep queries inside your perimeter.

Caveats

• The README is quickstart-heavy and leaves the underlying context-engineering mechanics largely unexplained.
• The “skills” plugin system and MCP integration are mentioned but not demonstrated, so their ergonomics are unclear.
• “Works with any data warehouse” is a broad claim, and the README never lists which connectors are actually implemented.

Verdict

Data teams that want to self-host a ChatGPT-for-their-warehouse and actually validate SQL correctness before users see it should look here. If you need a fully managed, turnkey BI solution with deep enterprise governance, this is still a build-it-yourself affair.