AWS simplifies AI agent OAuth with AgentCore consent portals

Overview

Amazon Web Services has introduced a managed consent portal for AgentCore Identity, a capability within Amazon Bedrock AgentCore. This new feature simplifies the complex process of managing OAuth grants for AI agents that act on behalf of end users. By offloading session binding and authentication infrastructure to a managed service, AWS aims to reduce the operational burden on developers deploying agent-based applications.

Previously, customers utilizing AgentCore Identity with an Authorization code grant were required to build and host their own session binding infrastructure. This involved presenting authorization URLs, hosting public HTTPS callbacks, authenticating returning users, managing browser sessions, and manually calling APIs to complete the flow. The new Consent portal replaces this custom engineering with a standardized web experience that handles these tasks automatically.

Streamlining OAuth Session Binding

The core innovation lies in how the Consent portal manages the relationship between a user’s identity and their application’s permissions. When an AI agent needs to access external services such as GitHub or Slack, it must first obtain explicit user approval through an OAuth flow. This process, known as session binding, ensures that the resulting access tokens are securely associated with the specific user who authorized them.

The managed portal provides a hosted web interface where users authenticate against their organization’s identity provider. Once logged in, they review the services available to the agent and grant consent for individual providers. The portal then handles browser redirects and session binding, storing the resulting tokens in the AgentCore Identity token vault.

This abstraction eliminates the need for organizations to maintain custom callback endpoints or manage fragile browser session states. Instead, the Consent portal acts as a secure intermediary that verifies user identity and permanently links OAuth grants to specific AgentCore Gateway targets. The administrator experience is similarly simplified, as configurations can be managed through the AWS Management Console with minimal boilerplate code.

Integration with IDEs and MCP Clients

One of the most significant use cases for this capability involves Model Context Protocol (MCP) clients and integrated development environments. Tools such as Kiro, Claude Code, Cursor, and Visual Studio Code increasingly rely on AI agents to interact with version control systems, communication platforms, and cloud resources. These agents require persistent, per-user access tokens to function effectively across multiple sessions.

With the Consent portal, developers can configure agents to request consent before invoking tools that require external permissions. For example, a software development assistant might need to list repositories on GitHub or post messages to Slack. Users can grant consent through the portal URL, and subsequent tool calls will automatically utilize the stored token vault entries. This ensures that access remains scoped to the individual developer without requiring repeated authentication prompts.

"This capability is particularly useful for agents accessed through IDE and MCP clients such as Kiro, Claude Code, Cursor, and Visual Studio Code. Users can grant consent before invoking a tool, and subsequent tool calls can use the token already stored for that user."

The architecture supports independent provider connections, allowing users to connect to GitHub without necessarily connecting to Slack. This granular control enhances security by adhering to the principle of least privilege, ensuring that agents only have access to the services explicitly approved by the end user.

Administrative Configuration and Security

Setting up the Consent portal requires a few prerequisite configurations, including an AWS account with Bedrock AgentCore access, a gateway configured with JWT inbound authorization, and administrative access to a corporate identity provider. The administrator must also register OAuth applications with providers like GitHub and Slack, ensuring the AgentCore Identity callback URLs are permitted.

During configuration, administrators define gateway targets for each service, such as a GitHub target for listing repositories and creating issues, and a Slack target for posting messages. An IAM execution role is created or assigned to allow the portal to discover these targets and manage credentials securely. The resulting portal URL is then shared with developers, who can initiate the consent flow directly from their browsers.

Security and auditability are maintained through integration with AWS CloudTrail. Every authentication event, consent grant, and token retrieval operation is logged, providing organizations with a clear audit trail. This visibility is critical for compliance monitoring, allowing security teams to review which users granted access to which services and when. The combination of managed infrastructure, granular permissions, and comprehensive logging positions this feature as a robust solution for enterprise AI agent deployment.