AWS’s November 17, 2026 walkthrough gives teams connecting AgentCore assistants to Snowflake a practical choice: individual identities or one shared service identity. Choose per-user OAuth—delegated authorization—when Snowflake needs to identify each employee. Choose a narrowly permitted service identity only when shared access and service-user query records are acceptable.
These are existing authentication options, not a new-feature launch. In this setup, AgentCore Gateway connects the assistant to Snowflake through Model Context Protocol (MCP), a standard interface for calling tools. The Snowflake-managed MCP server exposes services such as Cortex Agents, which can work with business data.
Individual sign-in is not, by itself, a guarantee of individual data filtering. Cortex Search uses the search service owner’s rights, so someone allowed to query an index can retrieve data they cannot read directly in its source tables. Check the tools’ permissions alongside the identity choice.
Which sign-in fits an employee assistant?
Use Snowflake OAuth when employees can authorize access through Snowflake itself. Its authorization-code flow creates a session for the user with an authorized role, without requiring an external authorization server. Plan for expiring access tokens: a valid refresh token can obtain replacements; without that capability, users must authorize again.
Use corporate sign-in when the assistant should use your existing identity provider. Snowflake External OAuth supports providers such as Okta and Microsoft Entra ID. User-delegated access can retain the employee’s Snowflake identity while using corporate authentication.
Corporate sign-in does not create the missing Snowflake permissions. The integration must validate the token and map its user claim to the intended Snowflake user. Check the issuer, audience—the service the token is intended for—signing keys and role grants. A successful login to the assistant is not proof that this downstream mapping is correct.
There is also an endpoint-level choice. Snowflake’s external-issuer binding rejects OAuth tokens from other issuers, including Snowflake OAuth tokens, at the affected MCP endpoints. The setting can apply at account, database or schema level. If different assistants need different issuers, plan those scopes before applying an account-wide default.
When is a shared service identity appropriate?
A shared identity fits an unattended application that genuinely needs one defined set of data permissions. In AWS’s client-credentials example, an external authorization server issues the token and Snowflake maps it to a service user. This avoids an employee consent flow, but it does not carry through the identity of whoever started the request.
For SQL queries, Snowflake’s QUERY_HISTORY records the executing user and active role. With the shared-service pattern, that user is the service account. Application logs can separately record who clicked a button; they do not turn the Snowflake query into a query executed as that employee.
The decision is therefore about acceptable access and accountability, not just convenience. If employees need different permissions, a shared credential cannot represent those differences by itself. Give a service identity only the access required for its application’s task.
Can a background task still represent a person?
Yes, when a valid upstream user token and a permitted delegation are available. AgentCore supports on-behalf-of token exchange: it exchanges an inbound user token for one intended for a downstream service, without another consent flow. The authorization server decides whether that delegation and the requested access are allowed.
This is another supported option, not proof that every unattended job needs a service account. Equally, it does not supply a human identity to a scheduler that has none. Check provider compatibility and token availability before choosing it. AWS’s Gateway compatibility table lists authorization-code, client-credentials and token-exchange grants for MCP-server targets.
Which permissions need a separate check?
The session’s actual role. Snowflake’s managed MCP server normally advertises
session:role:all. Despite its name, that selects the user’sDEFAULT_ROLE, not every role. Clients may differ in which role scopes they request. ADEFAULT_WAREHOUSEmust also be set for session initialization. Check the documented MCP role behavior.Extra roles and consent. For Snowflake OAuth,
OAUTH_USE_SECONDARY_ROLES = NONEprevents default secondary roles from being activated;IMPLICITactivates them. ALLOWED_ROLES_LIST restricts permitted roles and requires NONE. Do not confuse it withPRE_AUTHORIZED_ROLES_LIST, which skips explicit consent for listed roles. External OAuth has its own secondary-role controls.Each tool’s execution rights. Cortex Agents requires access to the agent and its underlying tools. Custom procedures may run with owner’s or caller’s rights. For Cortex Search, review who may query the service and what its index contains; employee authentication alone does not reproduce source-table restrictions.
There is a further boundary for customer-facing designs: every managed MCP request runs as a Snowflake user. Its agent arguments do not include the REST API’s per-request variables for identifying end users without Snowflake accounts. Snowflake directs that requirement to the REST API instead.
If the assistant also remembers conversations, handle that access separately. Our guide to per-user memory for a Bedrock document assistant explains why remembered context is not permission to retrieve documents.
What will the connection cost?
As of November 17, 2026, AWS lists AgentCore Identity at no additional charge when used through AgentCore Runtime or Gateway. Outside those paths, the listed rate is $0.010 per 1,000 token or API-key requests. There is no basis here for choosing shared credentials to avoid an assumed per-employee Identity fee.
Gateway API invocations are listed at $0.005 per 1,000. Illustration, not a workload estimate: 100,000 billable invocations × $0.005 ÷ 1,000 = $0.50 for that component alone. One question can involve several operations; an invocation is not necessarily one employee question. Gateway search and tool indexing have separate charges.
The larger budget still needs runtime, model inference and Snowflake consumption. Snowflake bills Cortex Agent orchestration and its tools separately: relevant drivers include orchestration and Analyst tokens, Search index size and persistence, and warehouse compute. Logs, network transfer and any identity-provider charges also need allowance. Without usage volumes and applicable contracts, an all-in monthly price would be speculation.
For another connector decision, our Apify–Notion permissions and costs guide separates credential handling, permitted operations and recurring charges.
What should the deployment owner verify?
Before enabling access, verify representative allowed and denied requests, the actual SQL user and role, and behavior when credentials expire or authorization is withdrawn. Include an owner-rights search tool in that review if the assistant uses one. These checks should establish the intended data boundary, not merely that sign-in succeeds.
Check the chosen deployment’s AgentCore component availability and Snowflake model and tool availability separately. AWS support for a Region does not establish that every part of a cross-vendor workflow is available there.
Choose the identity Snowflake should see first. Then configure the roles and tools to match that choice, and budget for the work the assistant actually performs.
