Add a support agent to your website or app that answers from your help centre and policies, looks up orders through your own API, and hands over to your team when it should.
Support teams that answer the same questions every day: delivery times, returns, refunds, password resets.
What it costs today
The answers already exist in the help centre and the policy documents, but customers don't find them, so they email or wait in a queue, and agents spend their day copying the same paragraphs.
What solved looks like
Customers get the right answer on the page they're on, with the source it came from, any time of day. People handle the cases that need a person.
Repeat questions answered straight from your own help centre and policies, with citations.
Order status looked up live, not guessed.
A handover message with your support address whenever the agent can't help or the customer asks for a person.
How it runs
Customer support on your website step by step
Each stage is a real step in SentientOne. Timings are medians from our latest benchmark; your model and network change them.
Customer support flow. A visitor asks on your website, through the widget with its publishable key. 1: the key is checked against your allowed websites and its limits (120 requests a minute per key, 12 per visitor). 2: the per-agent limit is checked and one credit is reserved. 3: your help centre and policy documents are searched with hybrid search and rerank. 4: the agent calls get_order on your orders MCP server. 5: your model writes the answer from the policy and the order status, and your prompt tells it when to hand over to a person. 6: the answer streams into the chat window with citations, and the conversation is on the agent's Conversations tab.
Channel
Website widget on your help pages (pk-so- key)
Visitor message
“My order 10482 hasn't arrived. Can I get a refund?”
Context
Allowed websites: northwind.example
Conversation kept until reload
X-Request-Id per message
SENTIENTONE
Widget key
Origin check
The pk-so- key only works from your allowed websites and only for this agent.
120/min per key · 12 per visitor
Credits, limits
1 credit
The agent's rate limit, then 1 credit reserved; refunded if the reply fails.
60 req/min · 402 at zero
Knowledge
~450 ms
Your crawled help centre and refund policy PDF, hybrid search and rerank.
help.northwind.example · refunds.pdf
MCP: get_order
Live data
The model looks the order up on your orders MCP server, behind a bearer token.
orders-mcp · 10 s per call
Generation
Grounded
Answers from the policy and the order. Your prompt says when to hand over to a person.
Claude Sonnet 4.6 on your key
Streaming
OTel trace
Markdown streams into the chat window, with a reference id if anything fails.
SSE · Conversations tab
Widget reply
Done · 1 credit
Order 10482 shipped on 21 October and is due on Monday. If it hasn't arrived by Wednesday you can have a full refund [1]. Want a person to look at it? Email support@northwind.example and quote your order number.
“My order 10482 hasn't arrived. Can I get a refund?”
Context
Allowed websites: northwind.example
Conversation kept until reload
X-Request-Id per message
message
SENTIENTONE
Widget key
Origin check
The pk-so- key only works from your allowed websites and only for this agent.
120/min per key · 12 per visitor
allowed
Credits, limits
1 credit
The agent's rate limit, then 1 credit reserved; refunded if the reply fails.
60 req/min · 402 at zero
credit
Knowledge
~450 ms
Your crawled help centre and refund policy PDF, hybrid search and rerank.
help.northwind.example · refunds.pdf
top 5 chunks
MCP: get_order
Live data
The model looks the order up on your orders MCP server, behind a bearer token.
orders-mcp · 10 s per call
order status
Generation
Grounded
Answers from the policy and the order. Your prompt says when to hand over to a person.
Claude Sonnet 4.6 on your key
tokens
Streaming
OTel trace
Markdown streams into the chat window, with a reference id if anything fails.
SSE · Conversations tab
agent loop · up to 8 rounds
first token ~2 s with a knowledge search and a tool round
answer
Widget reply
Done · 1 credit
Order 10482 shipped on 21 October and is due on Monday. If it hasn't arrived by Wednesday you can have a full refund [1]. Want a person to look at it? Email support@northwind.example and quote your order number.
Agents run on your own model provider key; there's no shared one, so nothing answers until you add it. Create a key with OpenAI, Anthropic, Google Gemini or Groq, then open Configure → LLM keys → Add key. It's encrypted at rest with AES-256-GCM, used only to call your provider, and never shown again in full. If a reply fails with INTERNAL_ERROR straight away, the agent has no key; with LLM_AUTH_ERROR, the provider rejected it.
Open Agents and choose New agent. Under Start from pick Customer Support. Give it a name, choose a provider and model, and pick the key you saved under LLM keys. A capable mid-size model writes clear support replies; pick a smaller one if speed matters most.
Replace the template's system prompt with your tone, what the agent may promise, and exactly when to hand over. Agents with knowledge also get a built-in rule to answer only from your content.
System prompt
You are the support assistant for Northwind Outfitters.Answer only from the knowledge base and the order tools. Quote the policy when you can.Before answering about an order, look it up with get_order. Never guess dates.If you can't answer, or the customer asks for a person, say so and givesupport@northwind.example with the order number to quote.Be brief and friendly. Use short paragraphs and lists.
On the agent's Knowledge tab: under Websites, crawl your help centre with max pages set to 40 and an include path like /help (200 credits). Under Documents, upload the refund and shipping policy PDFs (two PDFs, 10 pages, 50 credits). Under Q&A, add the 20 answers that must be word for word (20 credits).
On the Tools tab, choose Add server: your orders MCP server's URL, Auth set to Bearer with its token. Test connection lists its tools; you should see get_order. Use Refresh tools after you add tools to the server.
Open Playground, pick the agent and ask the questions your customers ask, including ones it shouldn't answer. Each reply shows its sources, tool calls, latency and tokens. Edit the draft prompt beside the chat until the answers are right.
On the Deploy tab, choose the style, colour, bot name and welcome line, add your domain under Allowed websites (example.com covers www; *.example.com covers subdomains), then copy the tag and paste it before </body>.
Your help pages
<scriptsrc="https://app.sentientone.ai/widget.js"data-widget-key="pk-so-…"data-agent-id="YOUR_AGENT_ID"data-style="bubble"data-color="#2563eb"data-bot-name="Northwind support"data-welcome="Hi! Ask me about orders, returns or delivery." defer></script>
Open Tracing to see every request: what knowledge it found, each tool call, tokens and timing, and the error code if it failed. Analytics charts requests, errors and response time per agent. Under Billing, turn on auto top-up so the agent never stops for want of credits.
Tell us about your setup and we'll help you build customer support on your website on SentientOne: the agent, the knowledge, the tools and the deployment.