Atlas vs Cube

Cube is the most mature open-source semantic layer, with ~19K GitHub stars and an enterprise-grade caching engine. Cube recently launched its D3 agentic analytics platform, adding AI-driven querying on top of the semantic layer. Atlas and Cube operate at different layers of the stack.

Quick Comparison

Different Layers of the Stack

Cube is infrastructure. It sits between your database and your application, providing a semantic layer with a powerful caching engine (pre-aggregations that materialize rollups). Applications query Cube via GraphQL, REST, or SQL APIs. The D3 platform adds AI agents that generate queries against the Cube semantic layer.

Atlas is product. It's an end-to-end agent that reads a semantic layer, writes SQL, validates it, runs it, and returns results with charts and narrative. It deploys as a standalone application, embeds in yours, or runs as a hosted SaaS at app.useatlas.dev with 3-region deployment (US, EU, APAC).

Trade-off: If you need a universal semantic layer that multiple tools query (BI dashboards, notebooks, AI agents), Cube is the right choice. If you need an AI agent that users interact with directly — embedded in your app or as a standalone tool — Atlas is purpose-built for that.

Semantic Layer

Both tools use YAML-based semantic definitions, but the scope differs.

Cube data models support complex transforms: pre-aggregations (materialized views for performance), multi-tenancy via security contexts, joins with fan-out control, and calculated measures with rolling windows. The Cube engine compiles these into optimized SQL.

Atlas entity files describe tables, columns, business terms, metrics, and query patterns. They're simpler — no compilation step — but they're auto-generated from your database via atlas init, editable through a web UI with schema-aware autocomplete and version history, and enrichable with LLM-powered descriptions. Atlas also has a dynamic learning layer: atlas learn proposes YAML improvements from query history, and runtime learned patterns are captured with admin review. Atlas includes its own LRU query cache (org-scoped, configurable TTL), though it's not a materialized rollup engine like Cube's.

Trade-off: Cube's semantic layer is more powerful for complex analytics workloads. Atlas's is simpler to bootstrap and maintain, especially when the primary consumer is an AI agent rather than a BI tool.

AI Capabilities

Cube D3 (launched June 2025) adds agentic analytics to the Cube platform with two specialized agents: the AI Data Analyst for self-serve natural language querying with visualizations and interactive data apps, and the AI Data Engineer for automating semantic model development from cloud data sources. Queries go through the semantic layer runtime rather than hitting the data warehouse directly — a trusted proxy architecture. Cube was recognized in the 2026 Gartner Market Guide for Agentic Analytics.

Atlas is agent-native from day one. The entire product is an agent loop: multi-step reasoning, tool use (SQL execution, Python analysis, semantic layer exploration), 7-layer SQL validation, and structured output. The agent doesn't just generate a query — it reads context, writes SQL, validates it, runs it, and explains the results. The backend uses Effect.ts with @effect/ai for typed, composable agent orchestration. Atlas also provides a notebook interface for multi-step exploratory analysis, a plugin marketplace with 21+ plugins, and 8 chat platform integrations via Chat SDK.

When to Choose Cube

• You need a universal semantic layer queried by multiple tools (BI, notebooks, AI)
• Pre-aggregation caching is critical for your query performance
• You want the most mature, battle-tested semantic layer in the ecosystem
• You're already using Cube and want to add AI querying via D3
• Apache 2.0 licensing is preferred (Cube core is Apache 2.0)