// the schema is the product

A semantic layerfor analytics,agent-native.

Atlas is a YAML-defined semantic layer for analytics — authored by humans, consumed by AI agents. Entities, glossary, and metrics live in your repo; the agent reads them, writes deterministic SQL, and runs it through a 7-stage validation pipeline — read-only, table-whitelisted, statement-timed.

Try the demo →$ bun create atlas-agent

self-host is free · MCP server for claude desktop, cursor, continue · slack-native

// the schema is the product

The YAML format is the moat.

Entities, dimensions, measures, joins, virtual dimensions, query patterns, glossary terms, and authoritative metrics — all in YAML, in your repo, code-reviewed in pull requests. Every field exists because an LLM needs it: sample_values ground the agent in real data, glossary.status: ambiguous forces clarifying questions, metrics.objective picks MAX vs MIN.

semantic/entities/orders.yml

name: Orderstype: fact_tabletable: ordersgrain: one row per order dimensions:  - name: status    sql: status    type: string    sample_values: [pending, shipped, delivered, cancelled]   - name: order_month    sql: TO_CHAR(created_at, 'YYYY-MM')    type: string    virtual: true measures:  - name: total_gmv_cents    sql: total_cents    type: sum joins:  - target_entity: Customers    relationship: many_to_one    join_columns: { from: customer_id, to: id }

atlas · agent replyvia MCP

// user

What's our top-performing category by GMV this month?

// agent reads orders.yml + categories.yml + glossary.yml, then writes SQL

SELECT c.name,
       SUM(o.total_cents) / 100.0 AS gmv,
       COUNT(DISTINCT o.id) AS orders
FROM orders o
JOIN order_items oi ON oi.order_id = o.id
JOIN products p ON p.name = oi.product_name
JOIN categories c ON c.id = p.category_id
WHERE o.status != 'cancelled'
AND o.created_at >= DATE_TRUNC('month', NOW())
GROUP BY c.name
ORDER BY gmv DESC
LIMIT 5;

// result · 5 rows · 7 validators passed

categorygmvorders

Bedding$184,2192,041

Kitchen$142,7181,587

Bath$98,4021,103

Outdoor$71,288812

Accessories$54,011693

// canonical questions

What's our GMV this quarter?

What's our top-performing category by GMV this month?

Monthly GMV trend over the past 6 months.

Show me revenue last quarter.

What are our most common return reasons?

// same questions on the readme, the docs homepage, and the eval harness · against the bundled NovaMart e-commerce demo

Read the YAML format →try the NovaMart demo

94%

of AI-generated SQL fails at least one Atlas validator.

// sample of 12,481 queries · gpt-4o, claude-sonnet, llama-3.1 · against 18 production schemas

// trace one query

One question, end to end.

Watch it run. Click any gate to see what it checks. This is the same panel the operator sees in the chat UI — every step is a real artifact, every gate is the real check.

session.4f8e · production.analytics

1.180s / 1.18s·7 / 7 gates·● ok

sqlplanresultrawresult

execute

Read-only connection, role-scoped session, results streamed back to the widget.

-- session.4f8e · 7 validations passed
-- read-only · scoped to analytics.public

SELECT c.name,
       SUM(o.total_cents) / 100.0 AS gmv,
       COUNT(DISTINCT o.id) AS orders
  FROM orders o
  JOIN order_items oi ON oi.order_id = o.id
  JOIN products p ON p.name = oi.product_name
  JOIN categories c ON c.id = p.category_id
 WHERE o.status != 'cancelled'
   AND o.created_at >= DATE_TRUNC('month', NOW())
 GROUP BY c.name
 ORDER BY gmv DESC LIMIT 5;

categorygmvorders

Bedding$184,2192,041

Kitchen$142,7181,587

Bath$98,4021,103

Outdoor$71,288812

Accessories$54,011693

step 11/11 · execute

// playbackautoplays on scroll · click any step or gate to inspect

// nodes in the system

Four primitives.

Inspectable, optional, TypeScript. No agent, no orchestration framework, no prompt salad.

entitysemantic_layer

Semantic layer

Entities, glossary, metrics in YAML. Versioned beside your code. Atlas reads them on every prompt.

entitiesglossarymetrics

gatevalidators

7 validators

AST-parsed, permission-checked, row-limited. Read-only by default. Same in dev, same in prod.

astpermsrow_limit+ 4

dbwarehouses

Warehouse-native

One connection spec. On self-host, no data leaves your network. Atlas runs in your VPC.

postgressnowflakebigqueryduckdb

auditaudit_log

Audit-ready

Every query, every result, every operator — logged, searchable, exportable. SSO, SAML, SCIM.

ssosamlscimcsv

// drop-in surfaces

React widget. Inherits your tokens, speaks your data.

dashboards.yml

Chat drawer is the editor. Per-user drafts, atomic Publish, persisted baseline.

prompt_lib.ts

Prompts in TypeScript, not strings in a UI. Diffed, rolled back, code-reviewed.

$ atlas cli

Run, test, replay queries from terminal or CI.

// why atlas

Three columns. One choice.

The same comparison from the README — same words, same scoring, no claim drift across surfaces.

feature

atlas

traditional bi

other text-to-sql

Semantic layer

Atlas: YAML on disk — query_patterns, virtual_dimensions, glossary.status: ambiguous, metrics.objective are all first-class

Traditional BI: Proprietary metadata, GUI-authored

Other text-to-SQL: None or limited

Agent-native

Atlas: MCP server first — Claude Desktop, Cursor, Continue with bunx @useatlas/mcp init

Traditional BI: Bolted-on AI feature

Other text-to-SQL: Standalone chat UI

Embeddable

Atlas: Script tag, React component, headless API, MCP, 8 chat platforms

Traditional BI: Standalone app

Other text-to-SQL: Standalone app

Dashboards as conversations

Atlas: Chat drawer is the editor — per-user drafts + atomic three-way-merge Publish over a persisted baseline

Traditional BI: Static dashboards with separate edit mode

Other text-to-SQL: No dashboards

Chat-platform reach

Atlas: Slack-native — answers questions in opt-in channels with a reaction-first tracer, fail-closed kill switch (Enterprise tier)

Traditional BI: Web app only — context-switch out of Slack

Other text-to-SQL: Web app only

Deploy anywhere

Atlas: Docker, Railway, Vercel, or your own infra

Traditional BI: Vendor-hosted

Other text-to-SQL: Vendor-hosted

Plugin ecosystem

Atlas: 21 plugins across 5 types — extend anything

Traditional BI: Closed

Other text-to-SQL: Limited

Open source

Atlas: AGPL-3.0 core, MIT client libs

Traditional BI: Proprietary

Other text-to-SQL: Varies

Multi-database

Atlas: PostgreSQL, MySQL, ClickHouse, Snowflake, DuckDB, BigQuery, Salesforce

Traditional BI: Usually one

Other text-to-SQL: Usually one

// deployment topology

Two ways to run it. Same code.

Cloud, or your VPC. Same Atlas, same primitives, same upgrade path.

// self-host

free

Your infra. Your data.

One command. Bun, Docker, or k8s. AGPL-3.0.
Every feature, no limits.

~/projects — bash

$bun create atlas-agent my-atlas
$cd my-atlas && bun run dev
→atlas booted on :3000
→connected · postgres://localhost
$▌

✓BYO model key
✓No telemetry
✓Community Discord

read the docs →

// atlas cloud

$29/ seat

Hosted. Zero ops.

recommended

We run it. Weekly updates, monitored connections, SLA.
Live in 3 minutes.

live status

atlas.openstatus.dev →

✓SSO · SAML · SCIM
✓99.9% uptime SLA
✓Audit log export
✓Priority support

start free trial →

// ship it

YAML in.Validated answers out.

Try the NovaMart demo →$ bun create atlas-agent

self-host is free · MCP server for claude desktop, cursor, continue · slack-native (enterprise)