The parts an LLM App is assembled from.
Last checked 2026-09-29
CX Enterprise atlas / Brand Visibility | AI & agents
Adobe LLM Apps
ProductBeta
Brand VisibilityAI & agents
Also known as LLM Apps, LLM Apps (Beta)
A beta application for building brand-owned apps that run inside AI assistants such as ChatGPT and Claude, for tasks like product discovery, availability checks and booking a service. Each app is a set of actions exposed as tools on one MCP server. Handler code runs on Adobe I/O Runtime from a GitHub repo, interactive widgets are Edge Delivery Services blocks, and analytics come from Customer Journey Analytics. The code lives either in repositories Adobe creates and hosts or in two GitHub repositories of the customer's own.
Integrates with 2
Integrated by 1
Contains 4 parts, 14 entries
- Actions# Capability
Also known as LLM actions
Brand-defined business flows, such as Compare Products or Book a Service, that the assistant invokes when a request matches the action's description. Metadata, input schema, annotations and visibility are managed in the LLM Apps UI, and each action can be toggled on or off without a redeploy.
Last checked 2026-09-29
- Action handlers# Capability
Server-side functions in the customer's handler repository. They validate input, call brand APIs and return text for the LLM plus structured data for the widget.
Last checked 2026-09-29
- Widgets# Capability
Visual responses shown in the chat, such as product cards, carousels, booking forms or store locators. They are built as Edge Delivery Services blocks, governed through AEM, and published to an aem.page URL.
Integrates with: Edge Delivery Services
Last checked 2026-09-29
- LLM Apps SDK# API / SDK
Also known as @adobe/llmapps-sdk
The package that connects an Edge Delivery Services widget to the LLM host. Its embed script loads the widget page and hands each block a bridge for host styling, theme changes, calling other actions and resizing.
Last checked 2026-10-20
- LLM App MCP server# Service
The endpoint each deployed environment exposes, which LLM platforms use to find and call the app's actions.
Last checked 2026-09-29
- Automatic app creation# Capability
Analyses the brand's public website, proposes actions, and generates action metadata, handlers, tests and widgets for review, which can take up to 30 minutes. Works with either infrastructure option.
Last checked 2026-10-20
- Adobe managed infrastructure# Capability
The recommended setup, in which Adobe creates and hosts the action and widget repositories and connects the services an app needs, in about two minutes. Code access is available on request.
Last checked 2026-10-20
- Action from scratch# Capability
Defines new action metadata, writes the handler and connects a widget by hand.
Last checked 2026-09-29
- Your own repositories# Capability
Also known as Bring your own EDS project
Connects the app to the team's own GitHub repositories instead of Adobe managed infrastructure: two empty repositories that the platform initializes, or compatible existing action and Edge Delivery Services projects.
Last checked 2026-10-20
- Local handler development# Capability
Runs handlers as MCP tools on a local Node.js server for testing with curl or MCP Inspector.
Last checked 2026-09-29
- Deployment and testing# Capability
Builds and publishes the app to Adobe I/O Runtime in Stage or Production and keeps a history of recent deployments. The app can then be tested as a ChatGPT plugin in developer mode or as a Claude custom connector.
Last checked 2026-09-29
- LLM Apps analytics# Capability
Built-in reporting, powered by Customer Journey Analytics, on action triggers, success and abandonment rates, top user prompts and visibility scores. The analytics region is fixed when the app is created.
Last checked 2026-09-29