What happened
AWS’s announcement examines outbound authentication from Amazon Bedrock AgentCore Gateway to the Snowflake-managed MCP server, which exposes Cortex Search and Cortex Analyst as MCP tools. Gateway aggregates configured targets into a unified tool catalog. Amazon Bedrock AgentCore Identity brokers credentials: Gateway obtains a workload token, requests a resource OAuth token, and Identity manages the token vault, user-to-token binding, caching, and refresh.
AWS tested three configurations end to end. The first uses Snowflake as its own OAuth authorization server and the authorization-code grant, so each user signs in directly to Snowflake. The second uses authorization-code OAuth through a corporate identity provider such as Okta or Microsoft Entra ID; Snowflake validates the IdP-issued JWT through an External OAuth security integration. The third uses the client-credentials grant through an external IdP for a shared service user, with no user interaction or consent. AgentCore Identity supports custom and built-in credential providers, including CustomOauth2, OktaOauth2, and MicrosoftOauth2. AWS also notes API-key credential-provider support, including Snowflake Programmatic Access Tokens, but focuses on OAuth.
In the first two patterns, Snowflake can run queries as the real user and record that identity in QUERY_HISTORY; the two-legged pattern uses a shared Snowflake service user. The guide details controls needed to preserve those boundaries. For Snowflake OAuth, administrators can narrow advertised role scopes with OAUTH_SCOPES_SUPPORTED and restrict consentable roles with ALLOWED_ROLES_LIST. AWS warns that secondary roles can exceed the consented primary role if OAUTH_USE_SECONDARY_ROLES is set to IMPLICIT, and recommends NONE to avoid that expansion. For external OAuth, administrators configure claim mapping and validate issuer, audience, and signature. Some mappings are provider-specific: AWS gives sub for Okta and upn for Microsoft Entra as examples.
The source says access tokens default to a 600-second lifetime, with a maximum of 3,600 seconds. AgentCore Identity can use refresh tokens automatically; Snowflake’s refresh-token lifetime is controlled by OAUTH_REFRESH_TOKEN_VALIDITY, which the reference implementation sets to 86,400 seconds (one day), with Snowflake allowing up to 90 days. The guide says these patterns were tested with the Snowflake-managed MCP server using Okta and Microsoft Entra ID. The external-IdP patterns may extend to other OAuth-protected MCP servers, but token claims, scopes, and audience settings need validation for the target SaaS provider.
What the archive adds
New Runtime’s earlier coverage described an Okta Cross App Access governance pattern for controlling AI agent access with Amazon Bedrock AgentCore (7 November 2026). This AWS guide adds a distinct integration-level detail: how AgentCore Identity brokers OAuth credentials for Snowflake Cortex, including per-user and shared-service choices. It is a new fact on a related access-governance story.
Trend signal
This supports the pattern that agent security is moving toward runtime boundaries: AgentCore Identity brokers and refreshes credentials, while Snowflake’s security integrations, role scopes, and identity mappings determine what the agent can do. The prompt is not the control point described here. Agent security moves to runtime boundaries.
Why now
The guide addresses a concrete enterprise integration question: how to preserve existing user identity and authorization controls when agents access data through MCP. Its comparison is timely for teams connecting managed agent runtimes to corporate data systems and deciding whether queries need user-level accountability or a machine identity.
Connections to other events
New Runtime’s earlier coverage of Okta Cross App Access with AgentCore focused on governing agent access across applications (7 November 2026). This Snowflake guide meets that concern at the data-connector boundary: it details how OAuth grants, role scopes, and token identity shape what an agent’s calls can access.
Why it matters
The OAuth choice determines both operational convenience and accountability. A shared client-credentials identity can suit unattended jobs, but it does not provide the same per-user query attribution as authorization-code flows. Teams that need auditability should design the identity path before building agent behavior around the connector.
The guide also makes privilege boundaries tangible. A consented primary role is not necessarily sufficient if secondary roles remain active; IdP token claims must map to the intended Snowflake user; and issuer, audience, and scopes must match the target service. AgentCore Identity’s token vault and refresh handling can reduce custom credential plumbing, but they do not replace configuration review on the Snowflake side. For builders, the reusable lesson is to treat agent authentication as part of system architecture: select the grant based on workload identity, then test the actual authorization scope and audit trail.
What happens next
- More enterprise MCP integrations will document per-user OAuth and service-identity trade-offs. (6 months, 60%). Watch for: comparable vendor guides that specify identity mapping and audit behavior.
- Teams adopting these patterns will test role and claim configuration as part of connector rollout. (6 months, 55%). Watch for: published deployment checklists or implementation reports covering scope and audience validation.
Limits and uncertainty
AWS says the patterns were tested with Okta and Microsoft Entra ID against Snowflake’s managed MCP server; that does not establish behavior across every OAuth provider or SaaS MCP endpoint. AWS explicitly calls for validating token claims, scopes, and audience settings for other providers. The guide is an implementation account, not an independent security assessment, and it does not establish comparative operational costs or measured security outcomes. Teams still need to verify their own Snowflake role policy, IdP claims, token lifetimes, and audit requirements in their environment.
Skill to build now
Map one agent-to-data workflow before configuring it: identify whether calls must be attributable to an individual or can use a service identity, then document the grant, Snowflake role, token claims, and expected query-history identity. In a test account, confirm that the actual MCP call runs as that identity and that secondary roles cannot broaden access beyond the intended scope.
Reader takeaway
For an agent calling Snowflake Cortex over MCP, choose OAuth based on the identity the workload must carry: authorization-code flows for per-user accountability, client credentials for shared machine work. Then verify role scope, token claims, and audit behavior at runtime—not just the agent’s prompt.