Skip to main content

How to build context layer

Connect docs via MCP

A context layer is the governed layer that makes your existing data understandable and usable by AI at runtime. To build your first one, start with a single domain and a single source system that one team owns end to end. Connect that source to Atlan, enrich the missing metadata with context agents, have a domain expert review and certify the output, then expose the governed context to an AI client through the Atlan MCP server.

This guide covers the first single-domain rollout. The fastest path is one domain, one team, and one AI client, not a catalog-wide rollout on day one. For a multi-domain, enterprise-scale program, see How to build an enterprise context layer.

Prerequisites​

  • An Atlan workspace with admin access to install and configure the source connector.
  • Admin access to the Governance Center to run context agents in Context Agents Studio.
  • One source system that a single team owns end to end. Good first candidates include a Snowflake schema, a dbt project, or a BI asset collection.
  • A domain expert who can review AI-drafted descriptions and documentation.
  • A short list of 5 to 10 questions your team already asks of this data. You use these to validate the AI workflow later.
  • One MCP-compatible AI client to start with.

Pick one domain and one source​

Build a context layer one domain at a time. Trying to enrich the whole catalog at once is the most common way these projects stall.

  1. Choose a single source system that one team owns end to end.
  2. Confirm that the team has a clear domain owner who can review and approve the generated context later in this workflow.
  3. Write down the 5 to 10 business questions your team actually asks of this data today.
tip

Start with the source your team queries most often, not the largest one. Coverage matters less than relevance in the first iteration.

Connect source to Atlan​

Atlan connectors ingest the metadata, lineage, and usage signals that the context layer is built from.

  1. Open the connector setup for the source you picked.
  2. Authenticate using a method that the source supports and your security policy permits.
  3. Configure the source-specific settings needed for metadata, lineage, and usage signals.
  4. Run the first crawl and wait for metadata and lineage to populate.

For Snowflake, choose either the account usage method or the information schema method for metadata extraction. To enable SQL Intelligence later, also grant access to query history. For the full walkthrough, see Set up Snowflake.

note

SQL Intelligence requires query history from connected data warehouse sources. If the underlying SQL assets don't have query history, SQL Intelligence has nothing to analyze.

Enrich context with context agents​

Context agents are specialized AI agents in Context Agents Studio that enrich missing metadata using the context Atlan already has: table structure, lineage, query history, linked documentation, and glossary terms. They enrich only assets that are missing the target attribute, and never overwrite existing values.

  1. From the left sidebar in Atlan, click Governance → Context Agents Studio.
  2. Click the collection you want to enrich to open the collection detail view.
  3. Select the Overview tab to see which metadata attributes still have gaps and the coverage percentage for each.
  4. Run the Description agent first to generate business-meaningful descriptions for assets that don't already have them.
  5. Run the README agent next to generate higher-level documentation that explains purpose, usage patterns, and broader context.
  6. Run SQL Intelligence for tables and views that have associated query history from connected data warehouse sources.

To run an agent, find the attribute on the Overview tab, click Enrich Now, then click Generate & apply. For the full procedure, see Enrich metadata on asset collections.

note

Context Agents Studio currently ships the Description, README, and SQL Intelligence agents. Link Terms is coming soon. For what each agent generates, see Understand context agents.

Review, correct, and certify​

AI-drafted context is a starting point, not the final product. A domain expert still needs to review what ships.

  1. Review the highest-traffic assets first, not the entire collection at once.
  2. Accept high-confidence descriptions, edit weak ones, and add the business context that AI can't infer reliably, such as ownership nuances, edge cases, and metric caveats.
  3. Set a certificate status, such as Verified, Draft, or Deprecated, on the assets you trust so users can filter to trustworthy context.
tip

Don't try to certify every asset in the first pass. Most teams get the most value by certifying the smaller slice of assets that answers most of their validation questions.

Expose context through MCP and validate​

The Atlan MCP server is the bridge between Atlan's governed metadata and external AI tools. The Remote MCP server is hosted by Atlan, enabled for all tenants, and uses the same authentication and authorization model already in place in Atlan.

  1. Pick one AI client to start with.
  2. Follow the setup guide for that client to connect it to the Atlan MCP server.
  3. Ask the client the same 5 to 10 questions you listed earlier.
  4. Capture failures, add the missing context at the source, then run the loop again.

Remote MCP documents support for clients including Cursor, Claude, ChatGPT, Gemini, VS Code, n8n, Windsurf, and Microsoft Copilot Studio. Authentication is client-specific: some clients support both OAuth and API key authentication, while others are more restrictive. For example, ChatGPT supports OAuth for its connector flow and doesn't support API key authentication there.

Use Set up Atlan MCP as the setup index, then follow the client-specific guide that matches your environment.

See also​