From the source
Amazon's guide walks through three OAuth configurations it tested end-to-end for connecting AI agents on Amazon Bedrock AgentCore to the Snowflake-managed MCP server.
- The first is Snowflake OAuth, where Snowflake acts as its own authorization server and each user authenticates directly through the Authorization Code grant with no external identity provider.
- The second is three-legged OAuth through a third-party IdP, where users sign in with corporate SSO and Snowflake trusts IdP-issued JWTs via its External OAuth security integration.
- The third is two-legged OAuth, where a shared service user authenticates through the client_credentials grant with no user interaction or consent, which the post says suits machine-to-machine workloads.
- AgentCore Identity acts as the credential broker, owning the token vault, refresh and user-to-token binding, and fetching, caching and refreshing tokens automatically.
- The patterns were tested with Okta and
Microsoft Entra ID, and the post notes that the external-IdP patterns follow standard OAuth flows and can extend to other OAuth-protected MCP servers if token claims, scopes and audience settings are validated.






