Skip to content

MCP Servers

LMA’s Strands agent supports dynamic loading of MCP (Model Context Protocol) servers, enabling integration with external tools and services. MCP extends the meeting assistant’s capabilities beyond its built-in tools, allowing it to interact with CRMs, scheduling platforms, code repositories, and more during live meetings.

The Model Context Protocol (MCP) is an open protocol for connecting AI assistants to external tools and data sources. Rather than hard-coding every integration, MCP provides a standardized way for AI systems to discover and invoke tools at runtime.

LMA supports MCP servers that provide tools the Strands agent can call during meetings. When an MCP server is installed, its tools become available to the meeting assistant automatically. For example, installing a Salesforce MCP server lets the assistant look up customer records, while a scheduling MCP server enables booking meetings directly from the conversation.

LMA supports several authentication methods for connecting to MCP servers:

  • OAuth 2.1 with PKCE (recommended for web services) — The most secure option for browser-based OAuth flows. Uses Proof Key for Code Exchange to prevent authorization code interception.
  • OAuth 2.0 fallback — Standard OAuth 2.0 authorization code flow for services that do not yet support OAuth 2.1.
  • Bearer token — Simple token-based authentication for services that issue long-lived API tokens.
  • Custom header authentication — Allows specifying arbitrary authentication headers for services with non-standard auth schemes.
  • OAuth Client Credentials (machine-to-machine) — For server-to-server integrations that do not require user interaction.
  • Automatic token refresh — LMA automatically refreshes OAuth tokens before they expire, ensuring uninterrupted access to MCP server tools during long meetings.

The MCP Servers configuration page is available at /#/configuration/mcp-servers and is accessible to admin users.

The page contains two tabs:

Browse and install MCP servers from the public registry at modelcontextprotocol.io. The registry provides a curated list of MCP servers with descriptions, available tools, and installation instructions. You can search, install, update, and uninstall servers directly from this tab.

Configure custom MCP server endpoints that are not listed in the public registry. This is useful for internal or proprietary MCP servers deployed within your organization.

  • View the list of available tools provided by each installed MCP server
  • Install, update, and uninstall MCP servers
  • Maximum of 5 MCP servers per account

The following tools are available to the Strands agent without installing any additional MCP servers:

ToolDescription
list_meetingsList recent meetings
search_lma_meetingsSemantic search across meetings
get_meeting_summaryGet summary for a specific meeting
get_meeting_transcriptGet transcript for a specific meeting
start_meeting_nowLaunch a Virtual Participant into a meeting now (Zoom, Teams, Webex, Chime, Google Meet). Defaults to using stored Zoom credentials when present so the VP signs in to Zoom rather than joining as a guest.
get_virtual_participant_statusPoll the status of a VP launched via start_meeting_now. Returns the granular status (e.g. JOINING, MANUAL_ACTION_REQUIRED, ACTIVE, FAILED), a human-readable summary, the live VNC viewer URL, the meeting URL, and any errorMessage or manualActionMessage. Designed to be polled by the agent after a start_meeting_now call so it can verbalize “the VP is in the meeting” — or surface a CAPTCHA/2FA challenge to the user with the viewer URL — without further prompting.
schedule_meetingSchedule a future VP meeting

These built-in tools give the meeting assistant access to your meeting history and the ability to manage Virtual Participant sessions without any additional configuration.

When start_meeting_now returns a VP id, the agent should poll get_virtual_participant_status periodically until the status reaches ACTIVE (success) or FAILED (with errorMessage) or MANUAL_ACTION_REQUIRED (in which case the agent should surface the manualActionMessage and virtualParticipantUrl to the user so they can complete the challenge). See Virtual Participant › Status Lifecycle and Manual Action Required for the full state machine.

LMA uses a CloudFormation Custom Resource to automatically rebuild the MCP Lambda layer on stack create and update operations. This ensures that the correct native binaries are compiled for the Lambda execution environment.

Installed MCP servers and their configurations are preserved across stack updates, so you do not need to reinstall or reconfigure your MCP servers after updating LMA.

OAuth callback URLs are configurable for:

  • Quick Suite — Pre-configured callback URL for Amazon Quick Suite OAuth flows
  • Custom OAuth clients — Configurable callback URLs for third-party OAuth providers

These callback URLs are used during the OAuth authorization flow to redirect the user back to the LMA application after granting permissions.

LMA provides setup guides for several popular MCP server integrations:

  • Salesforce MCP Setup — Full CRUD operations on Salesforce objects (accounts, contacts, opportunities, leads, and more)
  • Amazon Quick MCP Setup — Connect LMA to Amazon Quick Suite (web, OAuth) or Quick Desktop (native, API key)
  • DeepWiki MCP Setup — Repository documentation search for accessing code documentation during meetings
  • Custom MCP servers — Install from the public registry or configure a custom endpoint

See also: Meeting Assistant