Skip to main content

Context layer

Not availableCoreAvailableStandardAvailablePlusAvailableProAvailableEnterprise Flex Compare
Private beta

The context layer is in private beta. Features may change. It's available to Airbyte Cloud organizations enrolled in the beta. Self-managed deployments of Airbyte aren't supported.

The context layer is how AI agents access and reason about your organization's data through the Airbyte MCP. It combines two ways to get data:

  • Replicated data: the data Airbyte syncs into your warehouse. It's complete, historical, and joined across all your sources.
  • Direct reads: live data your agent reads from a source or destination at the moment it needs it, without a sync.

What are direct reads?​

A direct read is a request your agent makes to a source or destination through the Airbyte MCP, outside of a sync. For example:

  • Your agent asks Zendesk for the current status of a ticket, instead of waiting for the next sync.
  • Your agent runs a SQL query on Snowflake to count last week's orders.

Direct reads use the credentials you already configured for that source or destination in Airbyte. You don't need to create a connection, and your agent never sees the credentials.

Direct reads are read-only. Your agent can list and get records, but it can't create, update, or delete them.

Why organization admins control the context layer​

Direct reads give AI agents access to data in your business systems. That's a decision for your organization, so only organization admins can turn on the context layer. Turning it on is a one-time task for your organization. After that, admins and workspace members with edit permission choose which sources and destinations agents can read from.

What an agent can read is also governed by:

  • The Airbyte permissions of the person or application using the Airbyte MCP. Airbyte users can only work in the organizations and workspaces they have access to, and can only perform actions allowed by their role in Airbyte.
  • The permissions of the credentials used to set up each source or destination. When an agent queries a source or destination directly, it uses the credentials saved in that connector in that workspace. This means it's possible for a user to see data they might not otherwise have access to, or to not see data they otherwise do have access to.

To turn on the context layer, see Manage context layer access.

How agents use the context layer​

After you turn on the context layer and give agents access to a source or destination, any agent connected to the Airbyte MCP can read from it.

  1. Your agent finds the source or destination with list_cloud_connectors.
  2. Your agent reads the connector's skill with get_agent_skill_docs to learn which entities and actions it supports.
  3. Your agent reads data with execute_external_api_query for a source, or execute_external_sql_query for a destination.

You don't need to name these tools. Ask your agent a question, like "What are the five most recent Salesforce opportunities?" and it picks the tools for you. For details, see Airbyte MCP tools.

Supported connectors​

Sources​

Direct reads are available for sources that are an agent connector. Agent connectors support the list and get actions on their entities. Direct reads don't support other actions that an agent connector documents.

When a source doesn't support direct reads, the Context layer page in Airbyte shows that it isn't supported yet.

Destinations​

Direct reads are available for these destinations:

Agents query these destinations with read-only SQL. Only SELECT statements and SHOW TABLES are allowed.

Usage limits​

During the private beta, the context layer is free to use. Each organization can make up to 100,000 direct reads per month. Other Airbyte MCP tools don't count toward this limit. Direct reads also count toward the API rate limits of the source you read from, and SQL queries use compute in your destination, which your destination provider may charge for.

Limitations​

  • The context layer is only available for Airbyte Cloud.
  • Direct reads are read-only. Agents can't write to sources or destinations through the context layer.
  • Direct reads are only available for supported connectors.
Was this page helpful?