New self-hosted analytics Project Strata Pitches a Semantic Layer That Can Refuse AI Queries

A new open project called Strata surfaced on Hacker News this week under the title "Show HN: Strata – an expressive semantic layer that can say no to your LLM," positioning itself as a self-service analytics platform built for an era in which both people and AI agents query company data.

According to its developer documentation, Strata is designed to turn raw database tables into what it calls a governed, business-ready semantic layer. The pitch is that domain experts can explore a data warehouse and export results to Excel or Google Sheets without writing SQL or filing a ticket with a data team. The project's tagline frames it as "high-performance, AI-safe self-service analytics," combining a governed semantic layer with dashboards designed to work for both human users and AI agents.

The documentation lists several core features: semantic modeling, where business-friendly fields, metrics and relationships are defined in YAML; automatic SQL generation with cost-based query routing; support for multiple data sources, including a data warehouse paired with faster OLAP tiers, with each query routed to whichever source can best satisfy it; version control for models through Git and migrations; and a REST API for metadata discovery and query execution.

Setup is built around Docker. A single command launches a bundled Strata server, which spins up its own PostgreSQL instance and generates encryption secrets on first boot, all persisted to a Docker volume. New users open a local web address, enter a free license key obtained from Strata's licensing site, and create an admin account. The server ships preloaded with a sample project modeled on the TPC-DS retail benchmark dataset, complete with a working model, sample reports and an AI agent to test immediately after signing in. The documentation notes that the bundled database is intended for local evaluation only, with a dedicated PostgreSQL instance recommended for production use.

Beyond the server, users install a command-line interface that requires Ruby 3.4.4 or later and Git. From there, developers can either clone the TPC-DS sample project, which runs on DuckDB and needs no external warehouse, or start a new project connected to their own data source. The documentation walks through cloning the sample, making a change, and deploying it back to the server with a single command, describing the whole process — from starting the server to running a live semantic model — as achievable in about 15 minutes.

The CLI supports a range of commands for managing the workflow, including adding and testing data source connections, listing tables in a source, creating semantic table models and relationship files, auditing a model for validity, and deploying changes to the server. Strata's documentation also points to further resources, including guides on core concepts, a hands-on tutorial extending the sample model, and a self-hosting guide covering production deployment, single sign-on and operations.

The available material, drawn from Strata's own product documentation and its Hacker News listing, does not independently verify every claim implied by the "can say no to your LLM" framing in the original post title, nor does it include performance benchmarks, adoption figures or third-party assessments. Readers interested in the project's real-world performance or its claims around AI query governance should consult the original Hacker News discussion and the project's documentation directly.