OPERATIONS

Support on your website, with a person one tap away

A contained agent answers your visitors around the clock from the knowledge you give it. It says up front that it is an AI, and when a visitor asks for a person, someone from your team replies in the same chat.

What you get

✓
Answers from your own knowledge

The agent answers from what you grant it: a knowledge folder, project documents and skills you approve. As a contained agent it sees nothing else on your server.

✓
Honest from the first line

Every website chat starts with “You're chatting with an AI assistant. A person can join if you ask.” Visitors always know who they are talking to.

✓
A person can join at any time

When a visitor asks for a person, the conversation shows up as waiting in your settings. Your reply appears in the same chat, labelled with your name and “(a person)”, and the agent stays quiet until you hand back.

✓
Costs you can predict

Four spending limits, checked before any cost: per answer, per person or visitor per day, for the agent per day, and for all contained agents on the server per month. By default they are $0.50, $2, $20 and $200, and you can change them.

✓
Visitor privacy built in

Visitors stay anonymous, the chat stores only part of their IP address, and conversations are deleted after 90 days by default. Visitor messages are processed by the AI model provider (Anthropic) to produce answers.

An agent answers, and a person can join at any time.

Visitors chat on your website at any hour. The agent answers what it can from your knowledge, and anything that needs a person goes to your team in the same chat.

Website chat · a conversation on harborcoffee.com
You're chatting with an AI assistant. A person can join if you ask.
> Visitor: Do you ship to Austria?
> Agent: Yes, we ship to Austria. Delivery takes 3 to 5 working days.
> Visitor: My last order arrived damaged. Can someone help?
> Agent: I'm sorry about that. I've asked a person from our team to join this chat.
Shown in your settings: waiting for a person
> Ana (a person): Hi, Ana here. I'm sending a replacement today.

✓ Answered from the shop's own knowledge
✓ Handed to a person when asked
A fictional shop. Your agent answers from the knowledge you give it.
HOW TO SET IT UP

For CRHQ clients

1

Turn on Contained

Create an agent for this job and, in its Contained settings, switch on Contained (the first time, press “Set up on this server”) and “Anyone can talk to it (visitors)”. Give it instructions for visitors and choose what it may read: a knowledge folder, project documents and skills you approve.

2

Set the limits

Under “On websites”, add the websites that may show the chat, then check the spending caps and how long conversations are kept. Contained agents run on a separate Anthropic API key you add for them.

3

Add the chat to your website

Paste one script tag into your site, or a plain iframe if you want the stricter option. The script tag runs on your page with your page's rights; the plain iframe runs nothing on your page.

4

Watch, join and improve

Follow the setup guide.

Read conversations as plain text in your settings and reply when someone asks for a person.

GOOD TO KNOW

Get more out of it, safely

Write its knowledge for the public

The agent only knows what you give it. Treat everything you give it as public: a determined visitor may get the agent to repeat it. Keep private details, passwords, keys and tokens out of its knowledge.

Start with small caps

The default caps keep costs predictable. Watch real conversations for a week before you raise them.

Allow only your own websites

The chat appears only on the websites you list, and links in replies are clickable only to those websites.

Tell visitors in your privacy notice

You decide what the chat is for, so your privacy notice should cover it and name the AI model provider (Anthropic) and Cloudflare Turnstile, which sees the visitor's IP address and browser information. You can delete any visitor's conversation at any time.

Put support on your site.
$299/month per server. Bring your own AI model. Deploy in 90 seconds.