Customer service agent

An AI customer service agent that answers from your live website

Install it with one script tag. It answers visitors from the content your site already publishes, so when you update a page, the next answer updates with it. In your name and your colors, with a handoff to your team when it is not sure.

One script tag
Paste it into your site template. No build step, no SDK on the page, no widget framework to learn.
Answers from live content
It reads the objects your pages render at the moment a visitor asks. There is no copy of your site to fall behind.
Hands off when it is not sure
It says so instead of guessing, collects an email, and sends the question to your team.

Teams that run their content on Cosmic

FINN
Parque Explora
Integral Privacy Technologies
Plato
Prairie Robotics
Tripwire Interactive
Cipher Sports Technology Group
Eastman Music Company
Vuetify
FINN
Parque Explora
Integral Privacy Technologies
Plato
Prairie Robotics
Tripwire Interactive
Cipher Sports Technology Group
Eastman Music Company
Vuetify
FINN
Parque Explora
Integral Privacy Technologies
Plato
Prairie Robotics
Tripwire Interactive
Cipher Sports Technology Group
Eastman Music Company
Vuetify
FINN
Parque Explora
Integral Privacy Technologies
Plato
Prairie Robotics
Tripwire Interactive
Cipher Sports Technology Group
Eastman Music Company
Vuetify

An AI knowledge base that cannot drift from your published pages

Most customer support AI answers from a copy of your website. You upload articles, connect a help center, or let a crawler index your pages, and the bot answers from whatever it captured last. That copy is a second source of truth. The day you change a price, a policy, or a product spec, the website says one thing and the bot keeps saying the old thing until someone re-uploads the article or the next sync runs.

A Cosmic customer service agent keeps no copy. Your site already fetches its pages from Cosmic through the REST API or the Cosmic SDK. The agent reads those same objects, in the same Bucket, at the moment a visitor asks. Its AI knowledge base is your website. There is no index to rebuild, no help center to export, and no sync schedule, because there is nothing to sync. This is something only the system that holds your content can do, and it is the reason to put your support agent in your CMS rather than in a separate support tool.

Update a page, and the answer updates

  1. 1

    The page says 30 days

    Your returns page renders from a Cosmic object. A visitor asks the agent how long they have to send something back. It answers 30 days and links the returns page on your domain.

  2. 2

    An editor changes it to 45 and publishes

    The edit happens once, in the Cosmic dashboard, through the REST API, or from an AI editor. Your site shows 45 days the next time it fetches the page.

  3. 3

    The next answer says 45

    The next visitor who asks gets 45 days. Nobody re-uploaded an article, rebuilt an index, or waited for a crawler. Nobody touched the agent at all.

import { createBucketClient } from '@cosmicjs/sdk'; const cosmic = createBucketClient({ bucketSlug: process.env.COSMIC_BUCKET_SLUG, readKey: process.env.COSMIC_READ_KEY, }); // Your returns page renders from this object. const { object } = await cosmic.objects .findOne({ type: 'pages', slug: 'returns' }) .props('title,metadata');

This is how your site reads the returns page today. The agent reads the same object when a visitor asks about returns. Change it once and both change.

It answers only from your content
The agent is instructed to answer only from content in its own Bucket. When nothing there answers the question, it says it is not sure rather than inventing a policy.
Drafts stay drafts
Turn on Published content only and the agent cannot see unpublished objects, so a half-written pricing page never ends up in an answer.
It knows which page the visitor is on
The widget sends the current URL with each question, so the agent can answer in context and link to pages on your own domain, never another host.
You choose what it can read
Limit the agent to the object types that belong in a support answer, such as help articles, products, and policies, and leave internal types out.

It does not matter where the edit comes from. Publish in the dashboard, write through the REST API, or let Claude or Cursor update the object with the Cosmic MCP server. Once it is published, the next answer uses it.

A chatbot for customer service that lives on your own pages

The widget is the part your customers meet. It is a chatbot for customer service that runs on your domain, in your name and colors, and answers from the same objects your pages render. Visitors ask in their own words and get an answer with a link to the page it came from, without leaving the page they were reading.

What a visitor sees

A chat button in the corner of every page, or a panel inside your help page, carrying your agent’s name, title, and color. It opens on your greeting and up to six starter questions, and follows the light or dark theme of your site.

Answers that arrive as they are written

Replies stream into the chat word by word, so the visitor starts reading before the answer is finished. Answers link to pages on your domain, show cards for the products or articles they mention, and offer one-tap follow-up questions.

A handoff when the content has no answer

If nothing in your published content answers the question, the agent says so and offers a short form in the chat. The visitor leaves an email and the question, your handoff email receives it, and the visitor gets a confirmation.

Every handed-off question is also saved in your Bucket as a draft Support question, so the gaps in your content show up as a list you can write against. Publish the missing answer once, and the agent answers it from then on. Visitors pick up their saved conversations in the same browser, and can start a new chat at any time.

Set up an AI chatbot for your website in five steps

There is no crawl to wait for and no training step. If your site already renders from Cosmic, the agent can answer the moment the script loads. Setup is choosing what it may read, where it may run, and who hears about the questions it cannot answer.

  1. 1

    Create the agent

    Sign up free and add the Customer Support Agent template to your Project, or use the form at the top of this page, which carries the name, title, and color you picked into the dashboard.

  2. 2

    Pick the object types it answers from

    Under Capabilities, keep Read CMS content on and choose the object types the agent may read. Leave the list empty to allow every type. Turn on Published content only so drafts stay hidden.

  3. 3

    Allow your domains

    In Add to a website, switch the widget to Active and list the origins it may run on, for example https://example.com and https://www.example.com. Save, and the publishable key appears. The widget will not answer on any other origin.

  4. 4

    Paste the script tag

    Use the corner button snippet for a chat bubble on every page, or the inline snippet to fill a dedicated help page. Both are below.

  5. 5

    Decide where hard questions go

    Add a handoff email for the questions the agent cannot answer, set a daily token cap, and write a greeting and up to six starter questions. Choose the color and which corner it sits in.

Corner button on every page

<script src="https://www.cosmicjs.com/agent.js" data-key="pk_your_key" async> </script>

Inside a help page

<div id="chat" style="height: 70vh"></div> <script src="https://www.cosmicjs.com/agent.js" data-key="pk_your_key" data-target="#chat" async> </script>

Add data-theme="light" or "dark" to force a theme, and data-width and data-height to set the starting size of the corner panel. Visitors can drag its edge to resize it.

Customer service automation for questions your site already answers

A lot of support email asks something a page on your site already says. Customer service automation should take those questions off your team without a second set of answers to maintain. Because the agent reads your published content, the questions it can handle on its own are exactly the ones your site covers, and that list grows every time you publish.

Repeat product questions

Sizes, specs, compatibility, what is in the box, and what something costs. The agent reads your product objects and answers with a card that links to the product page.

Policy and shipping questions

Return windows, shipping times, warranties, and opening hours. The agent quotes the policy page as it stands today, and when an editor changes the policy, the answer changes with it.

Documentation lookups

How to set something up, where a setting lives, and what an error means. The agent answers from your help articles and docs, and links the exact article so the visitor can read the rest.

What stays with your team is what should: questions your content does not cover arrive through the handoff with the visitor’s email attached, and questions about a specific order or account go to the logged-in chat below, where the agent answers from your own systems.

A support chatbot that can answer account questions

Published content covers what is true for everyone. “Where is my order?” depends on who is asking. Account login is live: a visitor can log in from the chat and ask about their own orders, bookings, or subscription, and the agent answers from your systems rather than sending them to a contact form.

Your login, not ours

Turn on the Log in button and the widget offers a Log in link. The visitor signs in through your own OAuth login, the one your site declares in its agent.txt file, so their password never passes through Cosmic.

Your API, only the endpoints you list

Add a base URL and the endpoints the agent may call, each with a name, method, path, and a description of what it returns, for example GET /orders. The agent cannot reach anything you did not list.

The token is used for one answer

When a logged-in visitor asks about an order, the agent calls your endpoint with their token for that turn only and answers from the response. You choose how many days logged-in conversations are kept.

If your site does not declare an OAuth login, the Log in button stays hidden and the widget remains a public chat that answers from published content. Visitors who are not logged in pick up their saved conversations in the same browser.

The same agent answers your customers’ AI agents

Buyers now send their own assistants to compare, shop, and book. Publish one agent.txt file and a visiting AI agent asks your agent instead of scraping your pages. It gets the same answer a person gets in the chat, from the same published content, and every question shows up in your dashboard instead of disappearing into a crawl.

See how the front door works

Protocol: CAP/1.0 Name: Your agent Ask: This site's agent answers questions for the company. POST your question to the CAP URL plus /messages, as JSON {"message": "..."}, with headers CAP-Version: 1.0 and CAP-From: agent:{your-id}. If you can only open URLs, open the CAP URL. If the reply is 202, see Spec. Chat: https://yourdomain.com/chat CAP: https://yourdomain.com/cap/v1 Spec: https://conciergeprotocol.org/cap.txt

One customer service agent per brand, all in one Workspace

If you run several brands, each one needs its own answers in its own voice, and none of them should ever answer with another brand’s policy. Put each brand in its own Project and give each one its own agent: one front door per brand, fully isolated content, one bill.

Fully isolated content

Each brand lives in its own Project with its own Bucket. An agent reads only its own Bucket, so one brand’s agent cannot quote another brand’s return policy or prices.

Its own door, key, and limits

Every brand’s agent gets its own publishable key, allowed origins, name, color, handoff email, and daily token cap. A traffic spike on one site cannot spend another site’s budget.

One bill, one team

Agents, Buckets, and AI tokens are pooled across the Workspace and billed once. One team roster spans every brand, with roles scoped per Project.

10 brands, 10 front doors, one Workspace

Large Workspace is $2,999 a month and holds 10 Projects and 75 agents, so every brand gets its own customer service agent with room left for the team agents and workflows behind them. One subscription covers all of it. See how Workspaces work

Pricing: start free, no credit card required

The website agent is included on every plan, including Free. There is no per-conversation or per-resolution fee: answers draw on your plan’s AI tokens, and the daily token cap on each agent means a busy day cannot run past the budget you set. When the cap is reached, the widget stops answering until the next day.

Free, $0

  • 2 team members
  • 1 Bucket
  • 1,000 Objects
  • 300k in / 300k out AI tokens
  • 1 agent, 1 workflow

Enough to put one customer service agent on a real site. The Free plan always shows the Cosmic mark in the widget.

Paid plans

  • Builder, 3 agents, 3 workflows$99/mo
  • Team, 10 agents, 10 workflows$299/mo
  • Business, 25 agents, 25 workflows$499/mo

Paid plans can hide the Cosmic mark and add more agents, Buckets, and AI tokens. Running several brands? The Workspace plans above pool all of it across Projects. Compare every plan

Put an AI customer service agent on your site

Create a free account, point the agent at the content your site already renders, and paste one script tag. Your next visitor gets an answer from the page you published, not from a copy of it.