ChangoAgentic Data Platform
An agentic data platform built on an Iceberg-native lakehouse foundation.
Chango Key Features
A self-hosted AI data platform where agents and data live in the same system.
Iceberg-native Lakehouse
Chango isn't just Iceberg-compatible — it's built around Apache Iceberg as a first-class storage model from the ground up. Spark, Trino, and Flink share a single Iceberg catalog (Apache Polaris), with ACID transactions, schema evolution, and time travel built in. Data is stored in standard Iceberg format on open storage — no engine or vendor lock-in.
One Self-Hosted Platform
Object storage, streaming, batch + SQL, OLTP + vector + graph, workflow, and agents — installed, run, and observed from a single control plane.
AI as a First-Class Citizen
A multi-agent runtime and a Model Context Protocol server ship as built-in components. The Ontul MCP server exposes tools like ontul_search_metrics and ontul_describe_semantic_view for metric discovery, natural-language search (Korean 매출 ↔ revenue), and certification metadata — LLMs reach your data through a standard interface, not custom glue.
Ontul Semantic Layer — One Truth for Agents
Ontul, Chango's data engine, ships a production-grade semantic layer. Define a metric once and LLM agents, BI dashboards, and analysts all see the same number from the same definition. Aggregation, JOINs, and RBAC run server-side, so agents only need column names — answering with certified business definitions, not hallucinated formulas.
Native Retrieval Stack
Vector similarity, full-text search, and graph traversal live in the same engine as your transactional data. No external vector DB, no sync gymnastics.
One Identity Plane
A single IAM policy governs SQL queries, agent tool calls, MCP requests, object reads, and stream consumption — consistent across every layer. Add **single sign-on** — OIDC, SAML 2.0 or LDAP / Active Directory — and your existing identity provider becomes the source of identity: the groups it reports map onto platform groups, and permissions still come from the policies on those groups, so changing provider changes no policy. No local account is created, so removing someone from the directory removes their access.
Cross-Engine Lineage & Audit
Captured at the authorization layer — no separate lineage collector, no log-shipping agent. Every table access on Trino, Spark, and Flink becomes an audit event, including refused attempts with the reason they were refused, and every successful write (CTAS, INSERT, MERGE) becomes a source-to-target lineage edge. Reporting is asynchronous and best-effort, so it never slows a query down.
Open Engines Welcome
Every engine (Spark, Trino, Flink, Kafka) shares Apache Iceberg as the table format and Apache Polaris as the catalog — under the same IAM, the same observability, and the same operator experience.
CDC Without Leaving the Console — Kafka Connect & Topics
A Kafka cluster installed by Chango runs Kafka Connect too. Debezium connectors for PostgreSQL, MySQL, SQL Server and MongoDB ship inside the bundle, so change data capture installs with one click on an air-gapped host that can reach nothing; your own connectors upload as a zip, a tar.gz or a bare jar. The master keeps the authoritative copy of every plugin, so adding nodes later replays them onto the new workers — a task moving to a fresh worker never fails on a missing connector class. Each plugin gets its own directory, so two connectors cannot clash over a dependency, and a plugin change restarts only the Connect workers, never the brokers. Connector configuration is proxied to Connect rather than copied, so there is never a second answer to what is running. Topics are listed, created and deleted from the same screen — internal topics shown and marked rather than hidden, and refused for deletion.
Operations You Can Trust
Cluster topology, live metrics, KMS-encrypted state, version-pinned component lifecycle, and audit-ready dashboards in the box. The cluster master key rotates without downtime — the outgoing key stays accepted for reading — and control-plane backups upload to any S3-compatible target on a schedule, reporting which key an archive needs before a restore touches anything. Deleting a product cluster retains that product's master key, so data left behind in an external database or on a separate data disk stays readable.
Backups That Cover the Catalog
Managed PostgreSQL is backed up and restored to S3 alongside the control plane — because the Iceberg catalog (Polaris) keeps its metastore there. Losing that metastore loses no data at all in object storage, and every pointer into it. The dump is produced on the node holding the database and uploaded straight to S3 rather than routed through the master, with cron scheduling, retention, and one-click restore from the console.
Built for Air-Gapped Delivery
Nothing in the install reaches for an external repository. One bundle crosses the boundary on media and installs everything from the control plane to the components. It carries checksums at three levels — the media, its contents, and each product inside — plus two bills of materials: a product inventory with versions, digests, licences and upstream sources, and a CycloneDX SBOM of every library. Advisories name libraries rather than products, so a customer's security team can answer "are we affected?" without asking.
Use Cases
AI-Native Analytics
Business users ask in natural language; agents reach the data engine through the Ontul semantic layer and MCP, returning answers from certified metric definitions under the operator's IAM.
Retrieval-Augmented Applications
Vector, full-text, and graph relationships in one engine power production RAG and hybrid retrieval — no external vector DB to operate.
Sovereign Data Platform
Run the entire AI data stack on-prem or in your own cloud account — nothing leaves your perimeter.
Multi-Tenant Engine Mesh
Spark, Trino, Flink, and Kafka share infrastructure under one identity and resource-group control, with tenant data kept isolated.
The data platform for the agentic era
Agentic. Sovereign. One Control Plane.
Agents, MCP, vector & graph, engines, and storage — one self-hosted platform.