Kindly fill up the following to try out our sandbox experience. We will get back to you at the earliest.
How Decube's MCP server extends Snowflake CoCo and CoWork with cross-system lineage, data quality signals, and compliance-ready context for regulated enterprises
Updated on
Snowflake Summit 2026 (June 1-4, Moscone Center, San Francisco) delivered one thesis above all others: the agentic enterprise runs on governed context, not raw compute.
Snowflake rebranded Snowflake Intelligence as CoWork — a personal AI agent for knowledge workers — and officially named its coding agent CoCo (formerly Cortex Code). Both agents depend on a shared context runtime called Cortex Sense, which connects to data assets, business definitions, and operational knowledge to ground agent responses in enterprise reality.
Alongside these agents, Snowflake shipped Horizon Context — a governed semantic layer that activates metadata for AI, with semantic views, AI-generated documentation, and metadata connectors for Snowflake-managed and adjacent assets.
The architecture Snowflake described at Summit 2026 is an open one. Cortex Sense pulls context from registered MCP sources — not just Snowflake's own catalog. That openness is deliberate: Snowflake knows the regulated enterprise data estate spans dozens of systems, and governing context across all of them requires partners that specialize in exactly that problem.
Decube is one of those partners.
The Decube MCP server is a Model Context Protocol endpoint that exposes Decube's unified context layer — catalog metadata, column-level lineage, data quality scores, business glossary definitions, and observability signals — to any MCP-compatible AI agent, including Snowflake CoCo and CoWork.
Decube is a Trusted Data Context Platform built for regulated financial services. It unifies four capabilities that typically live in separate tools:
The MCP server exposes all four layers as queryable context, available to Snowflake agents at query time through Cortex Sense.
Snowflake Horizon Context is a strong foundation. It generates AI-powered descriptions from Snowflake table and column metadata, builds semantic views with governed business definitions, and connects to select external tools including Tableau, Power BI, and dbt. This handles the semantic layer for assets that live in — or are managed through — Snowflake.
Most regulated enterprises run data estates that extend far beyond any single platform. A loan risk model trained in Snowflake ML draws training data from a core banking system on Oracle, moves through a nightly ETL on Informatica, stages in Azure Data Lake, and lands in Snowflake after transformation. The business definitions, quality signals, and lineage that give that model meaning span every hop in that chain.
Decube's MCP server extends Snowflake's context layer to cover that full chain. It connects to 50+ source systems — Oracle, SAP, Databricks, Kafka, Informatica, Airflow, and more — and delivers three context capabilities that are especially critical for regulated enterprise workloads:
Cross-system lineage. Decube traces data column-by-column across every hop, from raw source to Snowflake consumption. When CoCo generates a metric, it can see the full derivation path, not just the Snowflake portion.
Live data quality signals. Decube monitors freshness, completeness, and validity at the source and attaches those signals to every cataloged asset. When a table has an active quality issue — a null spike, a referential integrity failure, a freshness breach — Decube surfaces that signal to Cortex Sense before an agent writes a query against it.
Compliance-ready lineage evidence. For regulated financial institutions, lineage is an audit requirement. Decube generates point-in-time lineage evidence exportable for regulatory reporting, and the MCP server makes that evidence available to Snowflake agents so every AI-generated output can be traced back to its certified source.
Together, Horizon Context and Decube cover the complete context stack a regulated enterprise needs to deploy Snowflake agents at scale.
Snowflake CoCo and CoWork both support MCP connectors. CoCo uses Cortex Sense to retrieve context automatically when generating SQL, building pipelines, or answering data questions. CoWork uses the same context runtime to ground knowledge worker queries in enterprise definitions.
Decube registers its MCP server as a context source in Cortex Sense. From that point, every CoCo session and every CoWork agent query can pull from Decube's context layer in real time.
Here is what a CoCo interaction looks like with Decube in the context stack:
A data analyst asks CoWork: "What was our NPL ratio for retail lending last quarter, broken down by product?"
Without Decube in the stack, CoCo sees Snowflake schema names. It infers what "NPL ratio" means from naming conventions, picks what looks like the right table, and generates SQL. If that table had a quality issue on ingestion from the core banking system three weeks ago, CoCo has no way to know.
With Decube in the stack, CoCo receives the governed definition of "NPL ratio" from Decube's business glossary — certified by the risk team, version-controlled, mapped to the exact tables that implement it. It also receives the current quality score for those tables, the column-level lineage tracing the metric back to the core banking source, and any active observability alerts. CoCo generates SQL grounded in meaning and signals the analyst if any data quality conditions apply.
Cortex Sense, Snowflake's shared context runtime, was shown at Summit 2026 to improve query accuracy from under 50% to 83% when agents are grounded in rich business context. Decube extends that grounding to your entire data estate, so the accuracy improvement applies to every asset — not just those managed natively in Snowflake.
The clearest way to understand the joint architecture is to see each layer's job.
Snowflake Horizon Context governs the semantic layer for Snowflake-managed assets. It handles semantic views with governed metric definitions, AI-generated column and table descriptions, Semantic Studio for business logic editing, and Semantic View Autopilot for converting existing BI definitions into governed views. This is where the business meaning of Snowflake-native assets is defined and maintained.
Cortex Sense is the runtime that delivers that context to agents. It retrieves relevant context from registered sources — including external MCP servers like Decube — using hybrid keyword and semantic search, ranked by popularity signals and access control policies.
Decube's MCP server extends what Cortex Sense can retrieve. It contributes the context layer for the data estate outside Snowflake: cataloged assets from 50+ external sources, cross-system lineage, live data quality scores, observability incident context, and business glossary terms that span organizational boundaries.
The table describes a layered architecture, not a competition. Snowflake Horizon Context defines and governs the semantic layer for Snowflake-managed assets. Decube's MCP server extends that governed context to the rest of the enterprise data estate. Cortex Sense delivers both to every agent query.
CoCo generates SQL and builds pipelines from natural language. Its accuracy depends entirely on the context it receives at query time.
Cortex Sense with Horizon Context improved accuracy from under 50% to 83% for Snowflake-native queries at Summit 2026. That improvement came from grounding CoCo in governed business definitions rather than raw schema inference. Decube extends the same grounding to the cross-system context that the remaining queries depend on.
Three failure modes that disappear when Decube is in the context stack:
Semantic ambiguity. "Show me active customers" means different things in different business units. Decube's business glossary resolves the term to its certified definition before CoCo writes a single line of SQL — even when that definition was originally authored outside Snowflake.
Stale data. CoCo has no native signal for quality issues that originated upstream of Snowflake. Decube's observability layer tells CoCo which tables carry active quality flags at query time, so agents surface those signals to users rather than generating confident answers from compromised data.
Incomplete lineage. A derived metric is only trustworthy if its full derivation path is visible. Decube's column-level lineage gives CoCo the complete chain from source to Snowflake, enabling it to validate that a metric is correctly implemented — not just that the SQL is syntactically valid.
Connecting Decube's MCP server to Snowflake CoCo and CoWork takes four steps and under one day for a standard deployment.
In your Snowflake account, navigate to Cortex Sense configuration and add Decube's MCP server URL as a context source. Decube provides a dedicated MCP endpoint per tenant, authenticated via OAuth 2.0.
Decube automatically detects Snowflake tables and views you have already cataloged. For sources outside Snowflake — Oracle, Databricks, dbt, Informatica — confirm your existing Decube connectors are active and metadata is current.
In Decube's TrustyAI interface, mark the business terms you want available to CoCo and CoWork as "MCP-published." This puts certified definitions — owned, reviewed, and version-controlled — into the context stream that Cortex Sense queries at runtime.
Set your Decube data quality and observability alert thresholds for MCP delivery. When a monitored asset breaches a threshold, Decube writes an active-incident context object that Cortex Sense picks up within the next context refresh cycle (default: 5 minutes).
Once configured, no further integration work is needed. New Decube connectors, new catalog assets, and updated glossary terms automatically become available to CoCo and CoWork as you add them.
Snowflake Summit 2026's clearest message for APAC financial services: agentic AI is no longer a pilot technology. Every major bank evaluating Snowflake is now also evaluating whether its data estate is agent-ready.
For banks under OJK (Indonesia), BNM (Malaysia), MAS (Singapore), BSP (Philippines), and APRA (Australia), agent-ready carries a specific regulatory meaning:
Snowflake provides the AI compute, the agent surfaces, and the in-Snowflake semantic layer. Decube extends that stack with the cross-system governance layer that regulated institutions need before they can deploy agents in production.
Banks that are already running Snowflake for analytics and ML should treat the Decube MCP integration as the step that makes their Snowflake agent deployment audit-ready. The combination means any metric CoCo or CoWork produces carries a complete, traceable, exportable evidence chain — the kind regulators will ask for when AI audit season arrives.
Superbank Indonesia and MUFG are already running Decube as their enterprise data context layer. The MCP integration with Snowflake extends that investment to every agent they deploy on Cortex Sense going forward.
What is the Decube MCP server for Snowflake? The Decube MCP server is a Model Context Protocol endpoint that delivers governed data context — catalog metadata, column-level lineage, data quality scores, observability alerts, and business glossary definitions — to Snowflake AI agents including CoCo and CoWork. It registers as a context source in Cortex Sense, so agents receive cross-system enterprise context at query time alongside Snowflake's native Horizon Context.
How do Decube and Snowflake Horizon Context work together? Snowflake Horizon Context governs the semantic layer for Snowflake-managed assets: semantic views, governed metric definitions, and AI-generated documentation. Decube's MCP server extends the context layer to the rest of the enterprise data estate — cross-system lineage, data quality signals from external sources, observability alerts, and regulatory lineage evidence. The two operate as complementary layers in the same Cortex Sense context runtime.
Which Snowflake agents can use the Decube MCP server? Any Snowflake agent that consumes Cortex Sense context can receive Decube context, including Snowflake CoCo (the AI coding and SQL agent), Snowflake CoWork (the personal knowledge worker agent), and custom agents built on Snowflake's Cortex AI platform with Cortex Sense enabled.
How long does it take to connect Decube to Snowflake? A standard Decube MCP server integration with Snowflake Cortex Sense takes under one day. Registration of the MCP endpoint, data source mapping, business glossary publication, and quality alert configuration can all be completed by a data engineering team in a single sprint session. Once connected, new catalog assets and updated definitions flow to Snowflake agents automatically.
Why does Decube matter specifically for regulated financial services on Snowflake? Regulators in APAC — OJK, BNM, MAS, BSP, APRA — require that AI-generated outputs carry traceable lineage and evidenced data quality. Snowflake's native context layer handles in-Snowflake semantics well, but most bank data estates span Oracle core banking systems, SAP, and legacy ETL platforms. Decube provides point-in-time lineage evidence and quality certification across that full estate, surfaced to Snowflake agents via the MCP server so every agent output is audit-ready.
What is Cortex Sense and where does Decube fit in it? Cortex Sense is Snowflake's shared context runtime, announced at Summit 2026, that automatically retrieves relevant context for Snowflake AI agents from registered sources using hybrid keyword and semantic search. Decube registers as a context source in Cortex Sense via its MCP server endpoint, contributing cross-system lineage, quality signals, observability context, and business glossary definitions to the same runtime that serves Horizon Context.
Can I use Decube with Snowflake even if I am not yet using Horizon Context? Yes. Decube's MCP server registers directly in Cortex Sense and delivers governed context to CoCo and CoWork regardless of your Horizon Context adoption stage. Many regulated enterprises start with Decube as their primary context layer and adopt Horizon Context incrementally for Snowflake-native semantic view management. The two integrate naturally when both are active.
First step to AI Readiness
Table of Contents
Sneak peek from the data world.
_For%20light%20backgrounds.svg)
.svg){kind=small}
