“Can I pay with cryptocurrency?” was one of the questions our fictional Trailhead Support bot could not answer from its knowledge. In a recorded ChatGPT session, we asked HoverBot's connector to show that conversation. It returned the shopper's question, the bot's reply and the fact that the reply was a fallback. The assistant did not have to guess why the question appeared on the unanswered list.
Chatbase announced MCP events in ChatGPT. Its examples start work when a ticket arrives or a bot escalates, then ask ChatGPT to summarize or draft a reply. That is a useful step beyond a person asking for a report. It also makes the boundary worth checking: what can the connected assistant read, and what can it change?
What the HoverBot connector showed
Our recording used a reviewer workspace filled with invented support data. ChatGPT first asked the HoverBot connection for questions the bot had failed to answer in the previous seven days. It then requested the full conversation for the first result. The exchange below is copied from that recording. Trailhead Support is fictional.
ChatGPT user: Show me the full conversation for the first of those questions.
Connector result: Shopper: “Can I pay with cryptocurrency?”
Trailhead Support: “I'm sorry, I don't have information about cryptocurrency payment options.”
Connector result: The bot marked this as a fallback response.
A shortened extract from the recorded HoverBot connector demonstration. The full bot reply also offers general help. The workspace and its conversations are fictional.
The distinction between a fallback and a normal answer matters to whoever reviews the gap. A smooth sentence can sound complete even when the bot found no policy to support it. Here the connector carried the status alongside the words. A builder can make the same test with one unsupported payment or return question on a sandbox bot: ask for unanswered questions, open the original exchange, then compare the answer with the source material the bot was allowed to use.
The recording also asked for a weekly conversation comparison and for knowledge sources mentioning returns. Those reads show a second use: the assistant can help an operator locate the relevant source before changing an answer. The numbers in the recording belong to that invented reviewer workspace. They say nothing about live HoverBot customer traffic.
The write request was refused
At the end of the same session, the user asked ChatGPT to change the bot's greeting. The connected assistant replied that the connection gave it read access to the configuration but no action to update the greeting. That was the right result for this request. Today an owner may separately allow the connector to add or edit knowledge and reply to a visitor. Neither permission changes the bot's greeting. Check the permissions on the actual connection before treating any assistant as read only.
We leave knowledge edits and visitor replies off by default because the first job of a support connector is to make evidence inspectable. An assistant should say which conversation led it to a proposed fix. If an owner enables an edit, they should check the changed source before relying on the next answer. A visitor reply needs an explicit human decision.
OpenAI's connector guidance describes how a server exposes tools to ChatGPT and how workspace owners review available actions. The protocol can support writes. HoverBot exposes reads by default; the owner can separately enable knowledge edits and visitor replies. Ask a vendor to name the actions a connection exposes today. A general promise that it is “secure” does not answer whether the assistant can change a customer record or send a message.
Where Chatbase MCP events are ahead
Chatbase's November release describes a trigger when a new ticket, human escalation, customer reply or completed call occurs. ChatGPT can then prepare a priority or a draft without an operator remembering to ask. Our recorded HoverBot flow starts with the operator's question; it does not show an event subscription. Chatbase is ahead on that trigger. We have not run its event path ourselves, so we cannot judge its reliability in a particular support queue from the announcement.
The two designs share a practical test. A draft reply should point back to the original conversation and the policy it used. If the bot escalated because it lacked a source, the draft should say that plainly. An event can make the work arrive sooner, but it cannot supply missing evidence. A team should decide who reviews the draft and which actions still need a person before connecting any event to customer-facing work.
ChatGPT and Claude are separate surfaces
HoverBot has a connection listed for both ChatGPT and Claude. Anthropic describes MCP connections for Claude; OpenAI documents its own setup and workspace controls. The recorded exchange above is from ChatGPT. It is not a Claude test result, and we would not claim the two assistants phrase the answer the same way. The useful common design starts with support reads: bot information, conversations, unanswered questions, knowledge and summary data. Check each connection for optional write permissions.
For an agency installing this for a client, the screen is only the last step. Start with a sandbox workspace whose questions and policies you know. Connect the assistant with the intended account, inspect its available actions, then ask a question whose correct result you can check without trusting the assistant's wording. Remove access when the test is over. Repeat the read in the other assistant before promising parity across the two.
Try these checks this week
Put one invented, unsupported shopper question into a sandbox bot. Ask the connected assistant to find unanswered questions from that period.
Open the original conversation through the connection. Check that it shows the shopper's words, the bot's reply and whether the reply was a fallback.
Ask which source supports a proposed answer. If there is no source, write the policy in the bot's knowledge before expecting a confident reply.
Ask the assistant to change a greeting, then inspect the available knowledge-edit and visitor-reply permissions. Record who can enable each one.
If you use events, trigger one synthetic escalation and inspect the draft, the evidence it cites and the person who must review it.
For a bot answering from a fictional business, choose an industry demo and ask a question its pages cannot answer. Use the reply to start a connector check. Test edit permissions separately.
Top comments (0)