Cube.dev

Cube.dev builds a semantic layer and an agentic analytics platform: shared definitions of business metrics underpin queries, conversational analysis, workbooks, and dashboards. Co-founded by CEO Artyom Keydunov and CTO Pavel Tiunov, Cube moved from providing infrastructure for other developers’ analytics applications to shipping its own AI-driven experiences. In October 2025, it made Cube Agentic Analytics generally available and unified its semantic-layer and analytics offerings under the Cube name and one interface.[8][9]

The central mechanism is to have models query governed metric definitions rather than reconstruct business logic from raw warehouse tables. Cube’s Semantic SQL lets an agent combine existing measures into new calculations while Cube resolves their aggregation levels and applies the caller’s access rules. For example, an agent can divide revenue by order count without requiring a predefined revenue-per-order measure or calculating over intermediate rows in its context window.[11]

This architecture originated in an earlier Slack analytics bot. Its history explains how Cube’s metric-to-SQL infrastructure became a standalone product before the company introduced first-party agentic analytics.[1][8]

In his October 2023 interview, Keydunov recalled starting Statsbot in 2016 as a side project to bring statistics from sources such as Google Analytics, New Relic, and production databases into Slack. He published it on Reddit, and Slack later featured it in its application directory. The first natural-language implementation used regular expressions; Keydunov credited Pavel with exploring models beyond that approach. He described the resulting interface as unreliable for one-off requests and unable to sustain the follow-up dialogue needed to clarify a user’s question.[1]

Statsbot separated interpreting a question from translating a multidimensional query into SQL. Cube supplied the latter step: customers defined what their metrics and dimensions meant, and the framework mapped those definitions to their datasets. Keydunov recalled that users increasingly valued this infrastructure for applications beyond Statsbot. He and Pavel open-sourced Cube.js in March 2019; in a later recollection, he described discussing that decision with Pavel while playing ping-pong.[1][2][3]

The founders established Cube Dev, Inc. to resource the growing project. Its September 29, 2020 announcement reported a $6.2 million round led by Bain Capital Ventures, with participation from Eniac Ventures, Uncorrelated Ventures, Innovation Endeavors, Betaworks, Overtime.vc, Slack Fund, and others. At that point, Cube reported deployment to more than 70,000 servers, more than 100 GitHub contributors, almost 2,000 Slack community members, and nine full-time engineers.[2]

Cube’s 2020 commercialization plan paired an open-source core with Cube Cloud, then in closed beta. The company committed to releasing core features under Apache 2.0 for server code and MIT for the frontend, while the commercial service would help developers deploy and scale Cube applications. On July 19, 2021, Cube announced a $15.5 million Series A led by Decibel. By then it had released Cube Store, which the company described as improving latency and concurrency for analytical queries such as Top-K.[2][3]

Cube Cloud became generally available on October 5, 2021. Keydunov linked its development to more than two years of community feedback and production-support work with Pavel. The service handled API autoscaling, Cube Store worker scaling, schema compilation, cache warming, GitHub-triggered deployments, monitoring, and query tracing. It also managed pre-aggregations—aggregated, optimized representations of source data that queries can use instead of processing raw records.[4]

On August 23, 2023, Cube and LangChain introduced a semantic document loader for embedding Cube’s data-model metadata in a vector database. Their published demo loaded Cube views, stored embeddings in FAISS, retrieved entities relevant to a user’s question, and constructed an OpenAI prompt to generate SQL for Cube’s SQL API. Streamlit supplied the chat interface.[5]

Keydunov explained the intended benefit as reducing the complexity of the model’s task: requesting a metric grouped by a dimension can replace generating warehouse SQL with several joins and safeguards against incorrect aggregation. This requires upfront work to define and describe the data. It also leaves organizational disagreements about metrics to resolve. His prescription was to treat the semantic layer as a codebase, using version control and pull-request discussions to collaborate on definitions and manage revisions.[1]

In December 2023, Brian Bickell, David Jayatillake, and Michael Irvine reported a benchmark of Delphi plus Cube. It adapted an 11-question subset of an insurance benchmark developed by data.world researchers Juan F. Sequeda, Dean Allemang, and Bryon Jacob. The test used Snowflake data, a Cube semantic model, and Delphi’s Cube REST API integration; each question ran in a separate conversation, with five repetitions. The authors reported 100% correct answers for that test and credited Delphi’s prompt engineering and design choices. The result concerned the combined system and this question set.[6]

On June 6, 2024, Cube announced that it had completed $25 million in financing, including investment from strategic partner Databricks. Nnamdi Okike, co-founder and managing partner of 645 Ventures, joined as a board observer. Cube positioned its product as an independent, interoperable semantic layer between data sources and BI tools, spreadsheets, embedded analytics, and AI agents. It also described a recently released AI API using retrieval-augmented generation to supply relevant semantic context to models.[7]

At the June 2024 announcement, Cube reported over 10,000 developers in its Slack community, 17,000 GitHub stars, and threefold year-over-year growth in Cube Cloud customers and bookings. The commercial growth figures were relative measures, without absolute customer or bookings totals. In June 2025, the company reported thousands of organizations using its open-source and cloud products combined.[7][8]

Cube introduced D3, short for “data in cube,” on June 2, 2025. The preview brought together Analytics Chat, workbooks with SQL and Python, data apps, and AI-assisted semantic modeling for internal and embedded use. Agents queried semantic definitions and created ad hoc calculations through Semantic SQL; humans could review and modify those queries through a visual editor or SQL code. The architecture separated building the semantic model, generating queries, and creating visualizations, with safeguards including compilation checks, linters, prompts, documentation, and memory retrieval.[8][9]

The October 29, 2025 general release dropped the D3 name. Cube’s Semantic Model Agent could bootstrap models from warehouse tables or convert existing projects such as LookML files; workbook and dashboard agents supported analysis and presentation, while Analytics Chat could run multiple queries and summarize findings. Cube retained its headless APIs and added embeddable frontend experiences. It also introduced seat pricing for users accessing Cube directly while retaining compute-based pricing for embedded analytics.[9]

In April 2026 research reported by Michael Rumiantsau, Cube compared schema-only prompts with prompts containing an additional 4 KB Markdown document of measures, conventions, and disambiguation rules. The benchmark comprised 100 retail questions over cleaned Contoso data in ClickHouse, crediting Bhanu Thakur for the cleaned re-export of Microsoft’s ContosoRetailDW. Cube reported accuracy rising from 50.5% to 67.7% for Claude Opus 4.7, 46.5% to 68.7% for Claude Sonnet 4.6, and 45.5% to 68.7% for GPT-5.4. This experiment tested supplied prompt context, with no runtime or tool calls.[10]

In his September 3, 2026 engineering account, Keydunov said Cube had shifted its focus toward agentic analytics about a year earlier, retaining the semantic layer as its foundation. He described the surrounding harness as the rules and code governing context, tools, persistence, result checking, and permitted actions. Concrete choices included rejecting invalid constructions such as SUM(MEASURE(revenue)) before execution, returning targeted recovery hints, and preserving data-model changes as commits on development branches.[11]

Query responses returned the first 100 rows with a total row count, query ID, and explicit truncation notice, allowing agents to fetch more pages rather than mistake a sample for a complete result. Cube also separated executable metric definitions from Markdown business explanations, certified queries, retrieved company documents, and remembered corrections. Its searchDataModel tool refreshed and searched the model on demand, filtered candidates by caller permissions, and returned compact results. Keydunov described accepting the extra refresh cost to avoid silently stale model searches.[11]

On September 17, 2026, Cube described connecting Snowflake CoCo, formerly Cortex Code, through MCP or the Cube CLI. Teams could import Snowflake Semantic Views into Cube, then let CoCo discover metrics, query data, and build workbooks and dashboards. Direct searchDataModel and runQuery tools operated with the connected Cube user’s permissions and kept CoCo directing the analysis; a separate chat tool delegated questions to a Cube agent. Cube’s open-source agent skills supplied instructions for model exploration, querying, and content creation through the CLI.[12]