Travel Concierge ADK AI multi-agent published as Google Workspace add-on

This project integrates the Travel Concierge—an advanced, multi-agent AI from the Agent Development Kit (ADK) samples—into Google Workspace.

Showcase

Chat

Gmail, Calendar, Drive

Overview

While the ADK provides default interfaces for developers (chat and text), this project focuses on creating a Graphical User Interface (GUI) dedicated to end-users. It demonstrates how to publish the Travel Concierge as a single Google Workspace add-on, making it accessible directly within applications like Google Chat, Gmail, Calendar, and Drive.

This solution serves as a practical guide and a reusable framework for developers who want to connect powerful, custom AI agents to Google Workspace, allowing users to leverage advanced AI capabilities seamlessly within their existing workflows.

Agent

The Travel Concierge ADK AI multi-agent being used is a conversational agent sample developed with the ADK. It's an advanced example, really close to a real-life AI agent. It uses tools like the Google Places API, Google Search Grounding, and a Model Context Protocol (MCP) server.

To expose this agent to end-users, a GUI with specific features is necessary as discussed here. This sample shows how to build such a GUI by extending Google Workspace.

Architecture

This solution relies on the following technologies: Google Chat API, People API, Vertex AI Agent Engine API, Python, and Google Cloud Functions.

Features

User Sessions: Sessions are managed in Vertex AI. Each user shares one session across all Workspace apps. Users can manually reset their session to start a new conversation.
Rich Messaging: Users send text messages and receive responses with rich text and graphical card elements (carousels, images, buttons, links).
Error Handling: Features configurable retries and informs users of unexpected failures with graceful interactions.
Gmail Context: Users can include the current email's subject and body in their message to the agent.
Chat Streaming: Agent interactions are displayed in real-time as they are processed.
Chat Attachments: Users can send messages with recorded audio or file attachments for extra context.
User Context: The agent can optionally include user profile data, such as birthdays (available in Gmail, Calendar, Drive).
Switch to Chat: Access the Chat app DM in a single click from other apps (Gmail, Calendar, Drive) to access Chat-only features.

Prerequisites

Google Cloud Project with billing enabled.

Set up

Configure the Google Cloud project
1. Enable the Google Workspace Marketplace SDK then the Vertex AI, Cloud Functions, Chat, Places, and People APIs
2. Create a Service Account and grant the role Vertex AI User
3. Create a private key with type JSON. This will download the JSON file.
Setup, install, and deploy the LLM Auditor ADK AI Agent sample
1. Use Vertex AI
2. Use the same Google Cloud project
3. Use the location us-central1
4. Use the Vertex AI Agent Engine
Publish HTTP endpoint to Google Cloud Functions
1. Add Chat app credentials file credentials-chat.json to the project root directory
2. Create and deploy a new Cloud Function with the project source code
3. Configure the environment variables and redeploy
Publish Google Workspace add-on
1. Configure Google Chat API
2. Complete manifest and configure a new HTTP deployment with it

Customization

The core logic supports any ADK AI agent hosted in Vertex AI Agent Engine. Key customization points are:

main.py: Defines the main UI layouts and user interaction logic (Google Workspace event handlers). Example: Add support for Calendar event or Drive document context.
vertex_ai.py: Manages the streamed agent events, sessions, and error handling. Example: Enable multi-sessions for separate user conversations.
agent_handler.py: Implements abstracted functions for orchestrating operations. Example: Synchronize message history across all host applications.
google_workspace.py: Handles API interactions with other systems to gather context or take actions. Example: Add functions to retrieve details of a Calendar event.
travel_agent_ui_render.py: Controls how agent responses are displayed to end-users. Example: Design a new card to show a person's profile and avatar.