Android: Voice Mode hides connected app tools; switching to text exposes them in the same chat
ChatGPT app version
1.2026.223 (Android build 2622320)
Subscription
Plus
Platform
Android 16, Pixel 10 Pro
Issue
In an ongoing ChatGPT conversation, connected app tools can be unavailable to the assistant while the conversation is in Voice Mode, then become callable immediately after exiting Voice Mode and continuing the same conversation in text.
Observed example: Notion was connected and available to the account. During Voice Mode, the assistant repeatedly concluded that no Notion write tool / connector was available in the session. After exiting Voice Mode and sending a text message in the same chat, Notion search/fetch/create/update tools were exposed and worked normally.
OpenAI's Voice Mode documentation states that Voice does not yet support some tools and that custom actions are unavailable. Even if this behavior is intentional, the current UX/tool metadata makes the limitation look like a missing or disconnected app, leading the assistant to give an incorrect diagnosis and send the user toward unnecessary manual work.
Steps to reproduce
- Connect an app that exposes callable tools in ChatGPT (observed with Notion).
- Start or continue a conversation in Voice Mode.
- Ask ChatGPT to perform an action requiring that connected app, such as creating/updating Notion content.
- Observe that the assistant cannot see/call the connected app tools and may state that the connector is unavailable.
- Exit Voice Mode without starting a new conversation.
- Send a text message asking for the same action.
- Observe that the connected app tools become available immediately and work.
Expected behavior
One of the following would avoid the failure mode:
- Connected app tools work in Voice Mode; or
- Voice Mode explicitly exposes to the model/user that external app tools are disabled in the current modality, so the assistant can say 'switch to text to execute this' rather than incorrectly concluding the connector is absent.
Ideally, switching modalities within the same conversation should not silently change the tool surface without a clear capability signal.
Actual behavior
The tool surface changes silently between Voice and text. In Voice, the assistant interprets the missing tool as a session/account connector problem. In text, the tool is immediately available.
Why this matters
This breaks agentic workflows that begin by voice and then require execution in connected apps. It also causes misleading troubleshooting: the user may be told to reconnect an app or manually perform work even though simply switching to text restores the tool.
3 Comments
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
I don’t think #38848 is actually a duplicate of this issue.
#38848 is specifically about PDF/document grounding becoming inaccurate after switching to Voice Mode. This issue is about the connected-app/tool surface itself changing between Voice and text.
The important failure mode here is not simply that a particular connector is unsupported in Voice Mode. It is that the tools disappear silently, without giving the model enough capability/context information to distinguish:
That causes the model to make an incorrect diagnosis and recommend unnecessary troubleshooting or manual work. In the same conversation, switching back to text immediately exposes the tools again. switching back to voice makes the tool unavailable, and the model lacks context to be able to explain the difference or what to do. The expected behavior would be for the tool to be available globally, But in the meantime, it would be nice if the model had appropriate context and was able to suggest switching modes as a workaround.
So even if connected apps are intentionally unsupported in Voice today, I think the bug is the silent capability change and missing model-facing context around it. No PDF or document grounding is involved in the reproduction.
I have the same exact issue on iOS version with voice mode not seeing github plugin. Exiting voice mode makes it work perfectly.