Strata: A Semantic Layer That Refuses Invalid Queries Before Your LLM Runs Them
Strata is a full-stack analytics platform built around a governed semantic layer — one that validates queries and returns a refusal instead of a wrong number when a request doesn't fit the model. It's built by Ajo, who spent four years at Netflix working on self-service analytics for non-technical business users.
The core design bet is strict naming. Only one thing called Revenue can exist in a project. If revenue maps to multiple tables, query grain plus speed decides which table gets used. That constraint is what makes the rest of the system work.
Cross-domain blending without double counting
Because facts share conformed keys (e.g. a conformed Date key), Strata can join two fact domains — Orders and Support — into one blended result. Each fact is aggregated to the common grain and joined on conformed keys, never fact to fact. The team claims the system cannot double count, and agents get an explicit refusal rather than a wrong number.
Partition and aggregate-aware routing
Strata routes each query to the fastest engine that can answer it, falling back to the warehouse when filters fall outside the hot tier:
- Hot tier: ClickHouse or Druid
- Warm tier: Snowflake
- Cold tier: AWS Athena
You can load a portion of your data into a fast engine while keeping full history in Trino or something cheaper. The router picks the faster tier when the query can be resolved there.
Measure types and cohorts
Five measure types ship out of the box: Standard, Complex, Snapshot, Exclusion LOD, and Inclusion LOD. On top of that, users can build custom calculations that span fact domains, and segments that restrict a view or a single measure to a cohort — e.g. new customer revenue beside total revenue in one view.
Model as code
The whole model is two YAML file types plus the naming convention. The workflow: your coding agent drafts the YAML, strata audit validates it against the warehouse, and Git carries it to production. No view/explore/cube to pick first — every step is validated and previewed as you build.
Who it's for and how to try it
Strata is explicitly not a semantic layer your existing BI tool connects to — the team calls that a dead end. Instead it ships its own dashboards, subscriptions, scheduled reports, a Google Sheets add-on, row-level security, plus an API and MCP server for anything you build on top. It's in early beta, not open source, but the free tier covers up to 25 users and runs entirely on your machine via Docker.
📖 Read the full source: HN LLM Tools
👀 See Also

Headless OpenClaw Setup with Discord via Docker Scripts
A GitHub repository provides scripts to run OpenClaw with Discord in a headless Docker container, avoiding the TUI/WebUI. It includes a management script with commands like claw init, start, and stop, plus preconfigured support for OpenAI Responses API, Chromium, and various tools.

Relay CLI tool saves Claude session context when rate limited
Relay is a Rust CLI tool that reads Claude's .jsonl session transcripts from disk and creates full snapshots of your session, including conversation, tool calls, todos, git state, and errors. It generates context prompts to resume sessions after rate limits reset.

Jan Adds One-Click OpenClaw Installation with Jan-v3-Base Model Integration
Jan now supports one-click installation of OpenClaw with direct integration to the Jan-v3-base model, keeping all operations local and private on your computer.

Open-Sourced CLAUDE.md Keeps Claude Code Agents Productive for Hours, Not Looping
A single 70-line CLAUDE.md file stops Claude Code agents from drifting into narration and looping on fixes. Sessions go from 3-hour failures to full productive lifecycles.