Guide
Connect an AI assistant to your workspace
Ask Claude, ChatGPT, Grok or any assistant that speaks the Model Context Protocol about your chatbots without opening the dashboard, and answer customers who are waiting for a person.
In this guide
What you get
The question most worth asking about a support bot is what it could not answer. That question is tedious in a dashboard: you read transcripts until a pattern appears. An assistant with access to your workspace can read them all and tell you.
HoverBot exposes your workspace as an MCP server. Once it is connected, your assistant can look at your chatbots, your knowledge libraries, your conversations, the questions that went unanswered, your analytics and your captured leads, and answer questions about them in plain language. If you allow it, it can also reply to customers your chatbot handed to a person.
Read-only unless you say otherwise, and scoped to you
Every tool but one is read-only. The exception, reply_to_visitor, is off unless you tick “Also let it reply to customers” while connecting, and it can only answer a conversation your chatbot already handed to a person. No tool can edit a chatbot, change your knowledge, contact a lead or alter any setting. Which workspace it can see is fixed when you sign in, not by anything the assistant sends, so it can never reach another tenant's data.
Connect it
One server address, and a HoverBot sign-in. Your assistant handles the rest.
- 01
Add HoverBot to your assistant
In Claude or ChatGPT, install HoverBot from the connector or plugin directory. In Grok, open grok.com/connectors, choose New Connector, then Custom. Anywhere else, add a custom MCP connector. Where it asks for an address, use the one below.
- 02
Sign in to HoverBot
Your assistant sends you to HoverBot to sign in and choose what it may do. Tick "Also let it reply to customers" only if you want to answer customers from your assistant.
- 03
Ask it something
Your assistant lists the available tools once connected. Ask a question about your workspace and it picks the tool it needs.
Server address:
https://mcp.hoverbot.ai/mcpFor a script or a tool without sign-in support, create an access token on the Connectors page in your dashboard and send it as a bearer token. Treat it like a password: anyone holding it can read everything listed below.
What it can do
Eleven tools. Ten only read; one replies to customers, and only if you allowed it. Your assistant chooses between them; you do not have to name one.
| Tool | What it returns |
|---|---|
list_chatbots | Every chatbot in your workspace, with its type and publish status. |
get_chatbot | One chatbot's configuration: tone, system prompt, greeting, and the knowledge libraries attached to it. |
list_knowledge_sources | The files, web links, text blocks and question-and-answer pairs in your libraries, with their processing status. |
search_knowledge | Knowledge content containing particular words. A keyword match, not a semantic search. |
list_conversations | Recent conversations between shoppers and your chatbots, with the opening message and message count. |
get_conversation | The full transcript of one conversation, in order, with the retrieval confidence recorded for each turn. |
get_unanswered_questions | Questions your chatbots answered from weak grounding, and questions that produced no answer at all, each with its conversation id. |
get_analytics_summary | Conversation and message volume, active sessions, response rate and average response time, this week against last. |
list_leads | Contact details captured by your chatbots, with the chatbot that captured each one and its current status. |
list_escalations | Conversations a chatbot handed to a person that are not resolved yet, with their latest messages and whether anyone on your team has answered. |
reply_to_visitor | Sends your reply to the customer in one of those conversations, as you. Only when you allowed replying while connecting. |
On unanswered questions
get_unanswered_questions reports two different failures rather than merging them, because they call for different fixes:
- Low confidence means the bot answered, but from material it was not confident matched the question. Usually the answer exists and is worded differently from how shoppers ask for it.
- No response means the turn produced no answer at all. That is a different problem and worth looking at directly.
Each question carries its conversation id, so your assistant can pull the full exchange with get_conversation and show you what actually happened.
Customers waiting for you
When your chatbot hands a conversation to a person, that person can be you, inside the assistant you already use.
Ask “Which customers are waiting for a person?” and your assistant lists them with the conversation so far. Tell it what to say and it sends the reply into the customer's chat, as you. It is the same reply you would send from Live Activity, a Telegram alert or an escalation email.
An assistant that works on a schedule, such as a ChatGPT dot, can check for waiting customers every few minutes and tell you when someone needs you.
If you run your own OpenClaw assistant, HoverBot can also tell it straight away: connect it on the Channels page in your dashboard with its hooks address and hook token, and each handover is posted to it as it happens.
Things to ask
Ask in plain language. Your assistant works out which tools it needs.
- “What did shoppers ask about returns last week that we could not answer?”
- “Summarise the questions my bot got low confidence on this month, grouped by topic.”
- “Compare my conversation volume and response rate against last week.”
- “Show me the transcript for the conversation where someone asked about shipping to Iceland.”
- “Which knowledge sources mention the refund window, and what do they say?”
- “Which customers are waiting for a person right now?”
- “Tell the customer asking about order 4821 that the refund went out today.”
Access and revocation
Tokens are shown once, when you create them. HoverBot stores only a hash, so nobody, including us, can recover the original. If you lose one, revoke it and create another.
The Connectors page lists every token you have created, when it was created and when it was last used, so you can tell which are still in service before revoking anything. Revoked tokens stay in the list rather than disappearing, so the record of what existed remains readable.
Revocation takes effect on the next request. There is no cache to wait out and no session to expire.
Tokens do not expire on their own. That is why they cannot be created from a support session: if our support team is helping you with screen access, they cannot mint a credential that would outlive the session.
Limits
What this does not do yet, so you are not surprised by it.
No changes to your setup. Apart from replying to customers, there are no tools for changing anything. If your assistant finds a gap in your knowledge base, it can show you the gap; you make the edit in HoverBot.
Keyword search, not semantic search. search_knowledge matches words, so a query sharing no words with the source text will not find it. Your chatbot's own answering uses semantic retrieval; this tool does not.
Replies only where a person was asked for. Your assistant cannot take a conversation from your chatbot on its own. It can answer only conversations the chatbot already handed to a person.
Result sizes are capped. Listing tools return at most 50 rows per call and default to fewer, so an assistant asked about “everything” sees a recent slice rather than your whole history.
Questions, or something missing?
If there is a question you want to ask your workspace and no tool answers it, tell us. The tool surface is shaped by what people actually ask.
Get in touch