Context Platform for AI Agents

A data catalog indexes structured metadata about data assets — schemas, lineage, ownership, quality metrics — and delivers it to humans through a portal interface. A context platform unifies that metadata with the unstructured organizational knowledge that gives it meaning: runbooks, decision logs, business glossaries, and policies. The result is a unified context layer that serves both humans and AI agents. The distinction is architectural, not a checklist of features: A data catalog is a tool; a context platform is infrastructure. For a full breakdown of where the two diverge, see our post on context platform vs. data catalog.

DataHub uses an event-driven architecture that continuously syncs metadata from 100+ data sources as changes occur, rather than relying on scheduled batch ingestion. The DataHub MCP Server exposes the resulting context graph to any MCP-compatible AI tool — including Claude, Cursor, and Windsurf — through a standardized interface agents can query at machine speed. The Agent Context Kit provides additional SDKs and native integrations for platforms like Snowflake Intelligence, LangChain, and Google ADK. Agents receive not just schema, but lineage, ownership, quality signals, and linked documentation in a single governed response. For example, when a Notion page is updated, DataHub reflects that change in the context layer in near real-time. Every AI system querying it gets the latest version, not last week’s snapshot.

DataHub manages the three main layers of context. 

1. Technical context is the structural foundation: schemas, column types, table relationships, data lineage, transformation logic, ownership, and dependency graphs. 
2. Operational context is the behavioral layer: query patterns, usage frequency, freshness SLAs, quality scores, incident history, and performance metrics.
3. Business context is the human knowledge layer: business glossary definitions, domain classifications, governance policies, runbooks, decision logs, and the organizational knowledge that explains why data is structured the way it is — authored natively in DataHub or ingested from Notion and Confluence. 

All three layers are indexed in a single queryable context graph and served to both humans and AI agents from the same governed source of truth. For a deeper look at the distinction between technical metadata and business context, see our post on business context vs. technical metadata.

Context engineering is the practice of curating what goes into a single AI agent’s context window at inference time: selecting retrieved documents, managing memory, orchestrating tool calls, and optimizing what the model sees for a specific task. Context management is the enterprise-wide capability that makes context engineering possible at scale. Where context engineering asks “what does this agent need to know right now?”, context management asks “how does every agent across the organization access trusted, governed, consistent context?” Context engineering solves the context problem within a single application. Context management solves it across the enterprise. For a full breakdown of where the two disciplines diverge, see our post on context engineering vs. context management.

Context management and RAG operate at different layers of the stack. RAG is a retrieval pattern that runs on top of context management infrastructure. You still need it to bring context into the model’s context window. Context management is the enterprise infrastructure that ensures what RAG retrieves is discoverable, trustworthy, governed, and consistent across the organization. RAG is a technique. Context management is the foundation that makes the technique reliable at scale. For a deeper look, see our post on RAG vs. context management.

Agents pull context from DataHub on demand through the MCP Server, APIs, skills, and SDKs. DataHub does not push context to agents at query time; instead, it maintains a continuously synchronized context layer that agents query when they need it. Agents can also write back to the context graph — adding tags, updating descriptions, assigning ownership — as part of governed workflows. The pull design is intentional: agents need selective context based on what a user is asking. Pushing everything upfront would be noise. Agents pull exactly what they need, when they need it.

DataHub Core, the open-source project that powers the platform, is freely available and maintained by a community of 15,000+ members. DataHub Cloud is the enterprise-managed offering built on that foundation, adding managed infrastructure, enterprise governance, agent-ready delivery via the MCP Server and Agent Context Kit, and the scale and SLA guarantees production deployments require. Organizations can start with the open-source project and move to DataHub Cloud as their context management needs grow.

DataHub integrates at two layers.

1. On the data side, pre-built connectors ingest metadata from 100+ systems, including Snowflake, Databricks, dbt, Looker, Airflow, and more, alongside documentation from Notion and Confluence. 
2. On the agent side, the DataHub MCP Server connects to any MCP-compatible tool, including Claude, Cursor, and Windsurf, while the Agent Context Kit provides native integrations for LangChain, Snowflake Intelligence, and Google ADK. DataHub also exposes the context graph via REST and GraphQL APIs for teams with custom agent frameworks, so it fits into existing infrastructure rather than replacing it.

DataHub enforces fine-grained access controls at the context layer, before context reaches any agent. Policies govern which agents and service accounts can access which datasets, columns, business terms, and sensitivity-classified fields. The same policy engine that governs human access applies to every agent request. This means an agent not authorized to see PII-tagged columns cannot surface them, regardless of how the request is framed. Every context access is logged with provenance metadata, creating the audit trail compliance teams require under GDPR, HIPAA, and emerging AI regulations.

Agentic AI is moving from pilot to production, and the gap between the two is almost always an infrastructure gap, not a model gap. According to the State of Context Management Report 2026, 82% of IT and data leaders agree that agentic AI cannot reach production value without a context platform. The same report found 61% frequently delay AI initiatives due to lack of trusted data, and 57% duplicate AI efforts across departments because teams can’t find or agree on authoritative context. Prompt engineering and RAG alone don’t solve fragmentation, governance, freshness, or provenance — and those are the gaps that block production deployment.

Most context platforms give agents a path to your data. DataHub gives agents a reason to trust it. The platform is built around four purpose-built capabilities that work as a system: 

1. Context Ingestion unifies metadata from across your stack — Snowflake, dbt, Looker, Confluence, and more — into a single context graph automatically, eliminating the manual stitching that fragments context across teams. 
2. Context Intelligence continuously extracts semantic meaning from query logs, BI dashboards, and dbt projects to auto-generate context documents, so you’re not starting from a blank slate. 
3. Context Hub gives domain experts a structured, low-friction review queue where they confirm and refine AI-proposed context rather than create it from scratch. 
4. Context Activation delivers that validated context to every agent in your ecosystem — Snowflake Intelligence, Databricks Genie, Claude, Cursor, custom LangChain agents — through MCP, SDKs, and APIs from a single governed source. 

The result is consistent, deterministic answers across every agent, without rebuilding context engineering pipelines for each one. In production, DataHub customers have seen a 2x improvement in analytics agent accuracy and a 60x improvement in speed of answering analytical questions.