Create actions creating-actions

IMPORTANT
LLM Apps is available to all AEM Sites Cloud Service customers to develop and manage their own apps. Once metering begins, each app invocation will count toward the customer’s AEM Sites Content Requests. Adobe will announce the metering rate and start date before metering takes effect.

Actions are the capabilities your app offers inside an LLM conversation, such as finding products, checking availability, or booking a service. Each action’s handler performs the task, and an optional widget presents the result.

After infrastructure setup, choose how to create actions from the app page. Automatic generation is available with both LLM Apps managed infrastructure and your own repositories.

In this guide: Learn how to generate actions from your website or create them yourself. For automatic generation, start the process and review the results before deploying and testing the app.

Choose how to build actions

Open your app. If Adobe is still preparing its infrastructure, wait for setup to finish.

In the Actions section, How do you want actions to be built? offers two choices:

Choice
Use it when
Build actions automatically
You want the platform to analyze your website and generate suggested actions, handler code, tests, and widgets. This is the recommended starting point.
I’ll create actions myself
You want to define each action and supply its implementation yourself.

Choosing the infrastructure in the create-app wizard does not select either of these options for you.

Generate actions from a website generate-from-website

  1. Select Build actions automatically.
  2. In Your website, enter a public HTTPS website that represents your brand’s products, services, or customer tasks. This example uses https://frescopa.coffee/.
  3. Select Build my actions.

Actions — automatic generation selected with the public website entered before starting

Use a website you are authorized to analyze. Do not provide a site that exposes confidential or access-controlled information.

Selecting Build my actions starts generation; entering the URL alone does not. You do not need to recreate the app or reopen the infrastructure wizard.

Follow generation progress

The app page shows Actions are being generated. Select Go to actions, or open Actions from the navigation, to follow progress.

The UI estimates up to 30 minutes for generation. Leave the page and return later if needed.

Actions — generation is running, with a link to follow progress on the Actions page

During generation, the platform:

  1. Analyzes the website and identifies customer tasks for the app.
  2. Creates suggested action metadata, including descriptions and input parameters.
  3. Generates server-side handlers and tests in the action repository.
  4. Generates widgets in the EDS repository.
  5. Prepares the generated actions for review.

While the platform analyzes the website, the Actions page shows Discovering actions for your conversational experience. Suggested actions appear as the platform discovers them; their code and widget previews may still be generating. Wait for generation to finish before reviewing each action.

IMPORTANT
Generated handlers initially use representative sample data. Website analysis does not automatically connect your app to production APIs, live inventory, payment services, or booking systems.

If generation reports a failure, read the visible error and check the troubleshooting guidance. Do not approve an action simply to bypass a generation error.

Review the generated actions review-generated-actions

Review each suggested action before deployment. Confirm that it represents a useful task and that the generated behavior matches what you intend the app to do.

Actions — generated Frescopa actions are ready for review and have not been deployed

  1. On the Actions page, select Review for an action that is ready.
  2. Review the action’s description, input parameters, annotations, visibility, and generated handler.
  3. Review the widget configuration and preview, if the action has a widget.
  4. Inspect the generated code and tests. With LLM Apps managed infrastructure, the platform opens pull requests; with your own repositories, it pushes a branch to each repository. Check what is sample behavior and what must be replaced with real integrations.
  5. When the action is ready, select Mark as reviewed.
  6. Return to the action list and repeat for the remaining actions.

The saved-action editor separates Action details, LLM Widget details, Application Insights, and ChatGPT Extensions. Review the generated widget and handler changes in their respective tabs.

Generated action review — Branch pushed explains how the branch is merged, with the editor tabs and Mark as reviewed

Selecting Mark as reviewed merges the generated changes if no merge blockers remain: pull requests with LLM Apps managed infrastructure, branches with your own repositories. If an open pull request blocks review, resolve the blocker and retry. Do not approve an implementation with unresolved code-generation or merge failures.

With your own repositories, your team can also merge the branch with its own process, such as a pull or merge request, before selecting Mark as reviewed. If the platform cannot merge the branch itself, for example because the target branch is protected or the branch has conflicts, the review shows which branch to merge into which base. Merge it, then select Mark as reviewed again.

After you review all generated actions, the Actions page shows All actions are reviewed!. To return to the app, select Go to App page.

Review is not deployment. The actions are not available through your app’s MCP server until you deploy the app.

Create actions yourself manual-actions

If you prefer not to generate actions from a website:

  1. In the app page’s action-creation choice, select I’ll create actions myself.
  2. Select I’ll create my actions.
  3. On the Actions page, select Create Action.
  4. Follow Create an action from scratch to define the metadata, implement a handler, and connect a widget if needed.

The infrastructure remains the same. Choosing manual creation does not create handler code or connect your business systems for you.

Add manual actions after reviewing automatically generated actions as needed.

Deploy and test

After reviewing the generated work:

  1. Deploy the app to Stage.
  2. Test it in ChatGPT or Claude, using the staging MCP server URL.
  3. Test prompts that match each action and unrelated prompts that do not call for that action. Check both the text response and widget interactions.

Before making the app available to customers, customize the handlers to connect your APIs and customize the widgets. Test the app in Stage to confirm that it works as expected. Deploy to Production only after testing succeeds.

recommendation-more-help
llm-apps-help-main-toc