AI Meeting Assistant with AWS and Webex

Architecture overview

This AWS Machine Learning blog post describes a meeting prep and follow-up assistant built on two main pieces: Amazon Quick chat agents and Cisco Webex MCP servers. According to the post, Amazon Quick acts as the single conversational workspace, while the chat agent retrieves Webex context through Model Context Protocol (MCP) servers hosted by Cisco.

The post explains that Amazon Quick connects to remote MCP servers and registers the tools those servers expose as actions the assistant can invoke during a conversation. Three Cisco Webex MCP servers handle the integration work:

• Webex Meetings MCP — finds upcoming and previous meetings, and retrieves meeting status, AI-generated summaries, recordings, and transcripts.
• Vidcast MCP — searches Vidcast videos, retrieves AI-generated highlights and transcripts, and recommends related or trending videos.
• Webex Messaging MCP — searches spaces, retrieves messages and threads, and can optionally create a follow-up message or threaded reply.

The post notes that these are hosted endpoints (for example, https://mcp.webexapis.com/mcp/webex-meeting), so you do not host the servers yourself. It also mentions the chat agent can bring in context from enterprise sources like Amazon S3, Google Drive, SharePoint, and Confluence, and can use over 100 pre-built action connectors for systems such as Slack, Outlook, Jira, ServiceNow, and Salesforce.

Integration with Webex APIs

The integration runs through Webex OAuth 2.0. The post instructs you to create an OAuth integration in the Webex Developer Portal, either as one integration with combined scopes or as separate integrations per MCP server. It recommends separate integrations to make least-privilege reviews and troubleshooting easier.

Per the post, the Webex Meetings MCP requires scopes including spark:mcp, meeting:schedules_read, meeting:participants_read, meeting:summaries_read, meeting:recordings_read, and meeting:transcripts_read. A prerequisite step is that an organization admin enables the relevant MCP servers in Webex Control Hub under Apps > Agentic Apps, setting access to Allowed for the appropriate users. The post warns that if these servers remain blocked, the OAuth connection from Amazon Quick will fail during integration setup.

Generative AI meeting preparation techniques

The first use case in the post demonstrates orchestration across all three MCP servers from a single prompt. The example prompt asks the assistant to prepare the user for a recurring sync: locate the upcoming meeting, review prior related meetings, summarize key decisions and action items, pull recent Vidcast highlights, check Webex messages for unresolved follow-ups, and produce a short prep brief.

The post walks through the tool sequence: the agent starts with webex-list-meetings, then retrieves prior context with webex-get-meeting-summary, webex-list-transcripts, and webex-list-recordings. It searches Vidcast with vidcast-search-videos and vidcast-list-shared-with-me, extracts context using vidcast-get-video-highlights and vidcast-get-video-transcript, and can add recommendations. Finally, it uses webex-search-spaces, webex-search-messages, and webex-get-thread to find open follow-ups. Amazon Quick assembles the brief from the tool outputs.

Post-meeting summarization and action tracking

The second use case continues after the meeting ends. According to the post, the agent locates the meeting that just ended and retrieves its AI-generated summary; if that summary is incomplete, it uses transcripts for grounding. It then searches Vidcast for related updates, finds the relevant Webex space, and drafts a follow-up message. Notably, the post states the agent should ask before posting unless the user explicitly requests posting—a sensible human-in-the-loop guardrail for outbound actions.

Practical implementation considerations

The post lists clear prerequisites: an Amazon Quick subscription with permissions to create integrations and custom chat agents (admin access recommended for initial setup), a Webex organization with Meetings, Messaging, and Vidcast available, the MCP servers enabled in Control Hub, and valid OAuth credentials.

A recurring theme is access scoping. The post emphasizes that the signed-in Webex user must have access to the meetings, summaries, transcripts, recordings, Vidcasts, spaces, and messages the agent should retrieve—the assistant inherits that user's permissions rather than bypassing them. For teams evaluating this pattern, the separate-OAuth-integration recommendation and the post-before-confirm behavior are the details worth carrying into your own design. As always, treat the AWS post as the authoritative reference for current setup steps and scope requirements.