Troubleshooting troubleshooting

Last update: October 19, 2026

IMPORTANT

Adobe LLM Apps is currently in Beta.

Features, workflows, and UI shown here do not necessarily represent the final state of the product. To join the Beta, send an email to llm-apps-beta@adobe.com.

Start with the observable symptom. Do not share credentials, tokens, private MCP URLs, or sensitive handler results while troubleshooting.

App creation and action generation

Symptom

What to try

Adobe managed infrastructure setup fails

Read the error under Setup failed, then select Retry. If the page cannot check progress, select Check again; setup continues in the background

No AEM program is available

Ask your administrator for program access. If loading failed, select Retry loading programs

New repositories do not appear

Use Connect another repository to grant the Adobe LLM Apps GitHub App access, return to the wizard, and select Refresh organizations and repositories

Action or widget repository validation fails

Use a compatible existing project or an empty repository. Use separate repositories for actions and widgets, then select Validate repository again

AEM Code Sync is not detected

Install AEM Code Sync for the widget repository, add yourself as a site administrator on its confirmation page, return to the wizard, and select Validate AEM Code Sync installation

Creating the app did not generate actions

App creation only sets up infrastructure. On the app page, select Build actions automatically, enter Your website, and select Build my actions. See Create actions

Actions are still generating

The UI estimates up to 30 minutes for generation. Follow progress on the Actions page, or leave the page and return later

Action generation reports failure

Confirm the website is public over HTTPS and, for your own repositories, that repository access is still valid. Contact the Beta team with the visible error message if the problem persists

Actions and handlers

Symptom

What to try

Action is not invoked

Enable the app’s plugin or connector in the conversation. Confirm Expose to AI model in the action settings and the action’s Visibility switch on the app page are on. Improve the description and redeploy metadata changes

Empty or error response

Run npm test, then call the handler with MCP Inspector or curl. See Local handler development and testing

Handler works locally but not after deployment

Check the latest pushed commit, confirm the runtime configuration, and verify that the action code identifier matches actions/<code-identifier>/index.js

Generated action cannot be marked as reviewed

Confirm handler and widget generation succeeded. Check the generated pull requests for merge conflicts, reload the action, and select Mark as reviewed again

Widgets

Symptom

What to try

Widget does not render

Verify the Script URL, Widget URL, HTTPS, EDS publication, CSP domains, and CORS headers

Widget renders but shows no data

Call the handler with MCP Inspector and compare its structuredContent shape with the fields read from bridge.toolResult

Widget works in direct preview but not in ChatGPT

Direct preview might use sample data. Test the deployed handler result and verify the EDS origin is allowed by CORS and CSP

Browser request is blocked

Add only the required origin to the correct CSP field and redeploy

HTTP Headers Editor cannot save the configuration

Use the AEM Configuration Service or ask the EDS administrator to initialize the site headers configuration

Do not log complete bridge.toolResult values when they can contain personal or sensitive data.

Deployment

Symptom

What to try

Deployment fails during Preparing

Verify the handler repository is linked and your Adobe Developer Console access is still valid

Deployment fails during Build app

Run npm install, npm test, and npm run build locally. Fix dependency, syntax, or test failures and push the changes

Deployment succeeds but changes are missing

Confirm the expected commit was pushed and redeploy to the same environment

Action remains Not deployed

Deploy again after reviewing the action or changing its metadata

Authentication authentication

Applies when Enable authentication is on. See Authenticate end users with your own identity provider.

Symptom

What to try

Actions are still public after saving the settings

Deploy the app again to that environment. Authentication changes take effect on the next deploy

Settings look wrong after switching environments

Confirm the Workspace picker shows the environment you intended. Stage and Production are configured independently

Sign-in does not start

Confirm that you deployed the app after enabling authentication and that the action uses Required. An Optional action only prompts when its handler asks for sign-in

The platform sends the user to the wrong sign-in page

Verify Issuer exactly matches your identity provider’s issuer URL, and that it is reachable over public HTTPS

Sign-in succeeds but every call is still rejected

Confirm your identity provider issues tokens whose audience is the app’s MCP server URL for that environment, and that the token is a JWT signed with an asymmetric algorithm. See Token requirements

The identity provider never receives any traffic

Your provider’s discovery, authorization, and token endpoints must be reachable over public HTTPS. Check for a firewall, WAF, or IP allowlist in front of it — your app can be reachable while your provider is not

The identity provider refuses to issue a token for the requested audience

Some providers only issue tokens for a resource that is registered with them. Confirm the app’s MCP server URL is registered as a resource identifier in your identity provider

A newly added scope is not requested at sign-in

LLM platforms cache the app’s published metadata for a few minutes. Wait, then retry the sign-in

The user is asked to sign in again for a scope

The action requires a scope the token does not carry. Add the scope to the client’s grant in your identity provider, or remove it from the action

The user receives repeated sign-in prompts

A handler that returns extra.challengeAuth() on every call without checking the caller’s identity requests sign-in again after authentication. Request authentication only when the identity the action needs is missing

The sign-in is rejected before the user reaches your identity provider

The redirect URI the LLM platform sent is not registered with your authorization server. Some platforms issue a different one for each connector, so register the value shown on the connector setup screen

Save is blocked with an unsupported scope message

Add the scope to Scopes supported, or remove it from the action that requires it

A public action still prompts for sign-in in Claude

Check that the connector uses Sign in when needed, not Sign in now. For an older deployed app, update the runtime and redeploy. A mixed app on the current runtime leaves public actions callable anonymously

An Optional action returns sign-in text but no Claude sign-in prompt

extra.challengeAuth() is a ChatGPT tool-result challenge. For a portable sign-in gate, use a Required action so the runtime returns the HTTP challenge Claude expects

The app prompts for sign-in even though every action is None

Turn Enable authentication off and deploy. While it is on, the app advertises an authorization server even when no action is gated

Do not paste access tokens, claim sets, or identity provider client secrets into a support request.

ChatGPT plugins

Symptom

What to try

Plugin does not appear

Open chatgpt.com/plugins, confirm that your account or workspace allows custom MCP servers, and install or connect the resulting plugin

Plugin creation fails

Copy the full HTTPS MCP server URL again from Test the app and choose the authentication method matching the app. Use No Auth only for a public app, not as a workaround for failed OAuth

Plugin connects but cannot invoke actions

Confirm the plugin is attached to the chat, the actions are exposed to the model, and the latest version is deployed

Plugin uses the wrong environment

Edit or recreate the plugin with the intended Stage or Production MCP server URL

Plugin still shows old action metadata

Deploy the changes, select Refresh on the ChatGPT connection, then start a new conversation

If the issue persists, record the app name, environment, failed step, time, and visible error message before contacting the Beta team. Do not include secrets or sensitive customer data.

recommendation-more-help

llm-apps-help-main-toc