Vexa

What the Vexa MCP server does

The Vexa MCP server gives an AI client access to meeting capture workflows. Vexa bots join Google Meet, Microsoft Teams, and Zoom calls as participants, then produce transcripts with speaker attribution. Transcripts can be consumed while a meeting is in progress or retrieved after the call ends.

The service also supports an agent workspace in the self-hosted deployment. Captured meetings are compiled into Markdown files in a git-based workspace, allowing an agent to use meeting history as context and record selected results back into that workspace. The hosted service provides the meeting bot and transcription path; the agent plane is self-hosted only.

How it works

A request identifies the meeting platform and its native meeting identifier. The platform values documented by Vexa are google_meet, teams, zoom, and jitsi. For Google Meet, for example, the native identifier is the meeting code from the join URL.

After a bot is sent, transcript data can be read through the API. Live output is speaker-attributed and moves from draft text to confirmed text. The self-hosted API also provides an agent chat endpoint whose response streams as Server-Sent Events. Text arrives in message-delta frames, while commit frames indicate content recorded in the workspace.

The runtime launches bots and agents in isolated, temporary containers. Docker is the default backend, while process and Kubernetes backends are also available in the broader deployment. These runtime details matter when operating Vexa yourself, but an MCP client can use the exposed meeting and transcript capabilities without managing each workload directly.

Setup and configuration

The Vexa MCP server can be used through Vexa's hosted service without installing the repository. A hosted account supplies an API key, which is sent using the X-API-Key header. The documented bot endpoint is https://api.cloud.vexa.ai/bots.

For self-hosting, clone the repository and run make all. The project targets Ubuntu 24.04 for production and requires make plus Docker Engine 26 or newer. A Mac with Docker Desktop is supported for local evaluation. The command pulls release images, starts the Docker Compose stack, and prints an API key and local URLs, including the API gateway.

Transcription is required for normal bot creation. A deployment can use a Vexa account token or a self-hosted GPU transcription unit. Capture-only operation is available by setting transcribe_enabled to false for a spawn or by setting TRANSCRIBE_ENABLED=false for the deployment.

Tools and capabilities

The Vexa MCP server is suited to workflows that need to:

• Send a bot to a Google Meet, Teams, Zoom, or supported Jitsi meeting.
• Retrieve a transcript for a meeting by platform and native meeting ID.
• Consume speaker-labelled transcript updates during a call.
• Ask the self-hosted workspace agent questions about captured meetings.
• Stream agent responses and identify workspace commits.

Limitations and notes

Jitsi joining and capture are reported as offline-proven, but live validation is still pending. The hosted service does not include the sandboxed knowledge-agent plane; that capability requires self-hosting. Vexa's default bot flow expects a transcription service and returns HTTP 503 when the required STT setup is missing, unless capture-only mode is explicitly enabled.