AWS releases open-source MCP server to automate cross-account promotion of Amazon Quick agents

AWS has published the Quick Resource Migrator, a sample Model Context Protocol (MCP) server that promotes Amazon Quick resources from one AWS account to another in a single tool call. The server runs on Amazon Bedrock AgentCore runtime, and the code is available in the aws-samples repository. This is a reference implementation, not a new managed Amazon Quick feature.

The problem it targets

Amazon Quick is described by AWS as its agentic AI companion for work. Teams build chat agents, action connectors (Slack, Jira and others), knowledge bases, flows and spaces in a development account. According to AWS, there is no native one-click way to promote those resources to QA or production accounts. Teams currently rebuild each resource by hand: recreating agents with the same instructions and starter prompts, re-attaching connectors, re-granting permissions, and reprovisioning the S3 bucket, bucket policy and data source behind each knowledge base.

The migrator relies on the fact that these resources are managed through the Amazon Quick API, part of the Quick Sight API surface. The API covers create, read, update, delete and list operations, permissions included.

How it works

• Selection: Choose a resource type (agent, connector, knowledge base, flow or space), then select by id, by name, or all. Migrating a space brings its linked resources across.
• Upsert behavior: The server describes the target first, then creates or updates. Re-running a migration converges on the same state rather than creating duplicates.
• No deletes: The migrator never issues a delete against the target.
• Preview: A read-only mode reports what a run would create or update before it commits.
• Permissions: The server calls the relevant Describe*Permissions API on each source resource and replays the same actions in the target, remapping principals to registered users there.
• Backups: Before updating any existing target resource, it writes a versioned snapshot of the resource and its dependencies to a dedicated S3 backup bucket. If the backup fails, the update is aborted. A restore tool can roll a resource back to an earlier version.

What does and doesn't move

• Agents: Instructions, identity, tone, starter prompts and welcome message are recreated. Action connectors are re-attached and remapped to the target account.
• Action connectors: Configuration is recreated. Secret values are never read from the source. Connectors are created with placeholder credentials and must be re-authenticated in the target.
• Knowledge bases: The knowledge base is registered, its data source recreated and permissions copied. For S3-backed knowledge bases, the target bucket and bucket policy are provisioned. The documents themselves are not copied.
• Flows: Flow IDs differ across accounts, so flows are matched by name. A same-named flow is updated, otherwise a new one is created.
• Spaces: Spaces are recreated and re-linked to agents, connectors and knowledge bases, with ARNs remapped. Linked resources must be migrated first so the target ARNs resolve.

Architecture and security

The design uses three accounts. A central runner account hosts the MCP server on AgentCore runtime. The server uses AWS STS to assume a read-only role in the source account and a read-write role in the target account, so no long-lived credentials are stored. Callers authenticate with an Amazon Cognito JWT (client-credentials grant, scope invoke). The runtime can run in VPC network mode. Any MCP-compatible client can drive it, including Amazon Quick itself.

Pricing is not disclosed in the post. Costs would come from the underlying AWS services (AgentCore runtime, S3, Cognito, CloudWatch Logs), and the source excerpt does not give figures.

What this means

The post shows a gap in agent platforms: building an agent is easy, but moving it through dev, QA and production is not. AWS is filling that gap with sample code rather than a built-in promotion feature, so enterprises must deploy and maintain the migrator themselves.

The design choices are what regulated teams look for. These include additive-only operations, mandatory pre-update backups, a dry-run preview, and replayed permissions in place of hard-coded ones. Two limits remain. Connector credentials need manual re-authentication, and knowledge base documents are not copied, so a promotion is not a complete environment clone. The post also shows MCP and AgentCore being used for infrastructure operations as well as end-user agents.