Blab AI
Amazon Quick Extends Agentic Workflows with Model Context Protocol Integration
Amazon Quick now lets foundation models reach out to the outside world with the Model Context Protocol (MCP), turning the platform into a fully‑managed, secure gateway for external data.
The update, delivered through Amazon Bedrock AgentCore Runtime and AgentCore Gateway, gives models the ability to pull in files, query databases, and call APIs via a standardized, secure interface. The result is stateful, multi‑turn conversations that cut hallucinations and make agents more reliable.
MCP servers expose tools as REST endpoints that foundation models can call. Built with the MCP SDK, they run in containers listening on 0.0.0.0:8000/mcp and require OAuth 2.0 authentication, which AgentCore Identity handles. Amazon Quick connects to an MCP server through an AgentCore Gateway that performs inbound authentication (verifying the caller) and outbound authentication (authenticating the gateway when it calls the MCP server).
To deploy, developers start with a local MCP server written in Python. A minimal example uses the FastMCP library with the stateless_http=True flag, mandatory for AgentCore Runtime compatibility. The server is containerised with a Dockerfile generated by the Bedrock starter kit, which also creates a .bedrock_agentcore.yaml file for runtime settings and deployment parameters. Once configured, the project is deployed to AgentCore Runtime, turning the MCP server into a managed, serverless endpoint.
Next comes the AgentCore Gateway. An IAM role that the gateway assumes is defined with permissions that include the runtime ARN of the MCP server. Two Amazon Cognito user pools are set up: one for inbound authentication and one for outbound authentication. Each pool defines a resource server with a custom scope named “invoke.” The inbound pool supplies the client ID, client secret, and discovery URL that the gateway uses to validate JWTs from Amazon Quick, while the outbound pool supplies the same information for the gateway’s OAuth client, registered in AgentCore Identity.
With the gateway configured, its target is set to the MCP server’s runtime endpoint:
` https://bedrock-agentcore.us-east-1.amazonaws.com/runtimes/{encoded_agentcore_runtime_mcp_server_arn}/invocations?qualifier=DEFAULT `
The outbound authentication is linked to the OAuth client created earlier. Once the gateway and its target reach a Ready state, the MCP integration can be registered in Amazon Quick.
In Quick, a new connector of type “Model Context Protocol” is created. The connector’s endpoint is the gateway’s resource URL, and its authentication settings mirror the inbound Cognito pool: client ID, client secret, token URL, and authorization URL. Quick then synchronises the tools exposed by the MCP server; the sync completes when the connector’s action is in the Available state.
After registration, the MCP integration can be added to a chat agent or Flow in Quick. The agent can invoke MCP tools through the Actions integration, and the tools’ responses are returned to the agent’s conversation. Quick’s test action APIs can verify that the MCP tools are reachable and functioning.
The pattern promotes reusability. Teams can expose a single MCP server that hosts common tools and sub‑agents, and other Quick agents can consume those tools without building custom connectors for each use case. The architecture also enforces security boundaries: the MCP server controls its own resources, while the gateway handles token exchange and policy enforcement.
Amazon Quick’s MCP support follows the broader industry trend of standardising AI‑tool interfaces. The MCP standard, introduced by Anthropic in November 2024, has been adopted by major AI providers, including OpenAI and Google DeepMind. By integrating MCP with Bedrock AgentCore, AWS extends its enterprise AI portfolio and offers a fully managed, serverless path for deploying and securing external tools.
The integration is available to Quick users with an Author or higher subscription. AWS recommends that users review the prerequisites: an AWS account, IAM permissions for AgentCore, Cognito, and CloudWatch, and access to Bedrock Anthropic models. The deployment process is documented in AWS’s sample notebooks and GitHub repositories.
In short, Amazon Quick now delivers a turnkey solution for connecting foundation models to external services via MCP. The combination of AgentCore Runtime, Gateway, Cognito, and AgentCore Identity provides a secure, scalable, and reusable architecture that aligns with the MCP standard and the enterprise AI needs of modern organisations.
The update, delivered through Amazon Bedrock AgentCore Runtime and AgentCore Gateway, gives models the ability to pull in files, query databases, and call APIs via a standardized, secure interface. The result is stateful, multi‑turn conversations that cut hallucinations and make agents more reliable.
MCP servers expose tools as REST endpoints that foundation models can call. Built with the MCP SDK, they run in containers listening on 0.0.0.0:8000/mcp and require OAuth 2.0 authentication, which AgentCore Identity handles. Amazon Quick connects to an MCP server through an AgentCore Gateway that performs inbound authentication (verifying the caller) and outbound authentication (authenticating the gateway when it calls the MCP server).
To deploy, developers start with a local MCP server written in Python. A minimal example uses the FastMCP library with the stateless_http=True flag, mandatory for AgentCore Runtime compatibility. The server is containerised with a Dockerfile generated by the Bedrock starter kit, which also creates a .bedrock_agentcore.yaml file for runtime settings and deployment parameters. Once configured, the project is deployed to AgentCore Runtime, turning the MCP server into a managed, serverless endpoint.
Next comes the AgentCore Gateway. An IAM role that the gateway assumes is defined with permissions that include the runtime ARN of the MCP server. Two Amazon Cognito user pools are set up: one for inbound authentication and one for outbound authentication. Each pool defines a resource server with a custom scope named “invoke.” The inbound pool supplies the client ID, client secret, and discovery URL that the gateway uses to validate JWTs from Amazon Quick, while the outbound pool supplies the same information for the gateway’s OAuth client, registered in AgentCore Identity.
With the gateway configured, its target is set to the MCP server’s runtime endpoint:
` https://bedrock-agentcore.us-east-1.amazonaws.com/runtimes/{encoded_agentcore_runtime_mcp_server_arn}/invocations?qualifier=DEFAULT `
The outbound authentication is linked to the OAuth client created earlier. Once the gateway and its target reach a Ready state, the MCP integration can be registered in Amazon Quick.
In Quick, a new connector of type “Model Context Protocol” is created. The connector’s endpoint is the gateway’s resource URL, and its authentication settings mirror the inbound Cognito pool: client ID, client secret, token URL, and authorization URL. Quick then synchronises the tools exposed by the MCP server; the sync completes when the connector’s action is in the Available state.
After registration, the MCP integration can be added to a chat agent or Flow in Quick. The agent can invoke MCP tools through the Actions integration, and the tools’ responses are returned to the agent’s conversation. Quick’s test action APIs can verify that the MCP tools are reachable and functioning.
The pattern promotes reusability. Teams can expose a single MCP server that hosts common tools and sub‑agents, and other Quick agents can consume those tools without building custom connectors for each use case. The architecture also enforces security boundaries: the MCP server controls its own resources, while the gateway handles token exchange and policy enforcement.
Amazon Quick’s MCP support follows the broader industry trend of standardising AI‑tool interfaces. The MCP standard, introduced by Anthropic in November 2024, has been adopted by major AI providers, including OpenAI and Google DeepMind. By integrating MCP with Bedrock AgentCore, AWS extends its enterprise AI portfolio and offers a fully managed, serverless path for deploying and securing external tools.
The integration is available to Quick users with an Author or higher subscription. AWS recommends that users review the prerequisites: an AWS account, IAM permissions for AgentCore, Cognito, and CloudWatch, and access to Bedrock Anthropic models. The deployment process is documented in AWS’s sample notebooks and GitHub repositories.
In short, Amazon Quick now delivers a turnkey solution for connecting foundation models to external services via MCP. The combination of AgentCore Runtime, Gateway, Cognito, and AgentCore Identity provides a secure, scalable, and reusable architecture that aligns with the MCP standard and the enterprise AI needs of modern organisations.
