Answer the questions your visitors ask
Runner serves the chat widget, gives the agent a few excerpts from your product's own knowledge before each reply, marks it as AI and keeps human handoff on the same conversation record.
Built into Runner
- One script tag, limited by default to the domains you list
- It tells every visitor it is an AI
- One tap asks for a person, and once a teammate takes over the AI stops replying
Worth knowing
It runs on an AI account you connect and arrives switched off until you turn it on; live replies also wait on Runner's platform switch and a production sign-off for the widget. When a pricing, legal, security or similar answer has no source, or the model isn't confident, it says a teammate will follow up rather than guessing, and no audio ever reaches Runner.
Your agent backs pricing and security answers with your own docs, or hands the question on
It's 11pm, a visitor on your pricing page asks whether the team plan includes single sign-on, and nobody is awake. Before the agent writes a word, Runner searches your product's knowledge for the visitor's own question and hands the model a few short excerpts. That knowledge is the Markdown and text files from the repositories you connected.
Everyday conversational help needs no citation, and some topics carry a stricter rule. A claim about pricing, your roadmap, legal terms or security goes out only as a verified fact citing one of the excerpts retrieved for that reply, by its exact document and version. A citation to anything else is treated as stale or made up, and refused.
A refusal never trims the answer. One refused claim, a low-confidence answer, or a question that needs a person swaps the whole reply for one line: the agent has no confirmed answer yet, doesn't want to guess, and has flagged the conversation for a teammate.
Picture this
A ten-person dev-tools startup keeps its plan limits in a docs folder in the repository it already connected. A visitor asks about seat limits and gets an answer built from that file. Another asks about a security certification nobody wrote down, and gets the teammate line instead of a confident guess.
One tap puts a person in the conversation, and takes the model out of it
Every panel has a Talk to a human button in its header. Tapping it records a handoff on the conversation and tells your team to open the handoff queue, where a teammate replies from the same conversation record the agent was writing to.
Some questions don't wait for the tap. When the agent recognises a request for medical, legal or financial advice, or a possible self-harm concern, it opens a handoff on its own.
When a teammate takes the conversation over, the model is out of it. The check runs after the visitor's message is saved and before any model is called, so nothing the visitor types from then on reaches the model provider, and a visitor who writes again is told a teammate is handling it personally.
It ships switched off, and every switch is one you can name
A widget is created disabled. Once it's on, a chat starts by default only from a hostname on the widget's own list, checked in the database when the session begins, and a request with no origin at all is refused. A wrong or disabled key gets a script that quietly does nothing, the same for both, so nobody can probe which keys exist.
Replies run on a model account you connect, for one product or the whole organization, with no quiet fallback to a shared key, and real model calls sit behind an operator switch that ships off. A live widget key answers only once its production certification has passed. Until every one of those is in place, the agent generates nothing, and each visitor is told a teammate will follow up.
No chat session exists until a visitor opens the panel. The first thing the panel says is that they're chatting with an AI assistant.
Questions buyers ask
Can an AI chat widget hand off to a real person?
Yes. A Talk to a human button sits in every panel's header, and tapping it records a handoff your team picks up from the handoff queue. When a teammate takes the conversation over, the AI stops replying to it.
What does the chat agent say when it doesn't know the answer?
That it has no confirmed answer yet, doesn't want to guess, and has flagged the conversation for a teammate. The whole reply is replaced, never edited down.
Do I need my own AI account for the chat widget?
Yes. Replies run on a model account you connect, for one product or for the whole organization, and a missing key never quietly falls back to a shared one. With no model available the agent writes nothing, and the visitor is told a teammate will follow up.
Does the chat widget record a visitor's voice?
No. Where a visitor's browser offers dictation, the browser turns speech into text and only the text is sent. No audio reaches Runner.
Does the conversation survive a page reload?
No, a reload starts a new conversation. The session lives only inside the page, and nothing is written to a cookie or local storage to carry it across.
Use what already works
Connect your own account with the AI provider chat supports to generate replies — chat runs on one model, on your key. Runner supplies a few excerpts from your product's knowledge, a clear AI introduction and handoff; the model supplies the conversation.
How teams often buy this job
Usually bought as Intercom.
Intercom
Trademarks of their owners. No affiliation or endorsement.