What changed
AWS has introduced three secure OAuth patterns for integrating AI agents with Snowflake Cortex AI through Amazon Bedrock AgentCore. These patterns address the challenge of extending authentication and authorization controls into the agentic layer, offering options for enterprises to manage access without broad service account permissions or custom integrations.
What this means in practice
The new OAuth configurations allow AI agents running on Amazon Bedrock AgentCore to authenticate with the Snowflake-managed MCP server. The three tested patterns include Snowflake OAuth (per-user, no external Identity Provider), three-legged OAuth with a third-party Identity Provider (corporate SSO), and two-legged OAuth with a third-party Identity Provider (shared service user). Amazon Bedrock AgentCore Identity acts as the credential broker, managing token acquisition, refresh, and user-to-token binding. These patterns enable enterprises to choose the most suitable authentication method based on their specific needs, such as per-user access or machine-to-machine workflows.
Limitations and availability
The patterns tested include Snowflake-specific configurations and those using external Identity Providers like Okta and Microsoft Entra ID. Snowflake OAuth (3LO) is unique to Snowflake, relying on its built-in OAuth authorization server, while the other patterns follow standard OAuth flows and can extend to other OAuth-protected MCP servers. Enterprises must configure the appropriate security integrations on the Snowflake side and ensure token claims, scopes, and audience settings align with their SaaS provider. The solution architecture involves components such as Amazon Bedrock AgentCore Gateway, Identity, and the Snowflake-managed MCP server, with the outbound OAuth path consistent across all patterns, differing only in the authorization server and grant type used.