Mason R.
Reply to approveHigh risk
Virtual AI support employee for ecommerce brands
Paste one link. Your AI support agent starts today.
You type your store address. Brand, platform, languages, catalogue, policy limits and mailbox settings are all read for you.
What's your website?https://yourbrand.comStart readingReading starts in the background the moment you submit — you don't wait for it. Amazon sellers can paste a storefront or product link instead.What does your support desk revolve around?
Email · live chat · Amazon buyer messages, one card queueSupervise in Chinese, reply in the customer languageShadow mode first — nothing sends until you say so
One queue
Whatever channel it came from, it becomes a card in one queue
Email, live chat and Amazon buyer messages land in one queue as cards. Agree and it goes out; disagree and you say one line in Chinese — said once, it holds.
4 waiting
EmailHigh risk
Mason R. · K1 forks arrived scratched
Wants a replacement or $429 back
Order pulled · 2 policies cited · draft ships a fork plus a coupon
3 attachmentsYes, do thatNo
Live chatUrgent 02:47
Customer waiting · order already being picked
Can the delivery address still change?
Our policy does not cover this, so I stopped
Teach itLater
Amazon24h SLA
Buyer · Amazon US
Wants the size changed to XL first
Outbound gate all clear · no links, no emoji
Yes, send itNo
Knowledge12 times this week
A new knowledge entry to confirm
Tracking issued, no scan yet
Updates 24-48h after the label; write back if it does not
ConfirmNo
One link
Everyone else hands you a form. We just read your store.
Step one elsewhere is a long form. Those answers are already written on your own site and domain — we read them, and you just answer yes or no.
Brand and platform
The site title and page fingerprints give us your brand name and storefront platform.
Markets and languages
Language tags, hreflang and currency symbols pin down your markets and the customer language.
The whole catalogue
Every product on sale in your own categories, with images and prices; Shopify reads from a public endpoint.
Mailbox server settings
MX records identify the mail provider, so host, port and encryption fill themselves in.
Policy boundaries
Refund window, return shipping, duties and warranty are read off your policy pages for you to confirm.
Help centre and FAQ
Support subdomain, help centre and product pages come along as sources the AI can cite.
The one thing you type
Your support mailbox password — a secret is the only thing that cannot be derived. Provider, host, port and encryption all come back from your MX records.
A day in your AI employee's life
It works like an employee, not software waiting for clicks
Morning · Inbox takeover
Syncs the mailbox and moves support mail into the KefuAgents folder; the rest is untouched.
Daytime · Drafting
Reads the issue, looks up the order, cites audited knowledge, drafts in the customer language.
Any time · it raises a hand
When it is unsure it stops by itself and becomes a card on your phone.
Delegated · Auto-send
Delegated categories send themselves; risky ones still stop and wait for a person.
Evening · Daily report
Reports what it handled, what it sent alone, and what it held back.
Your role
You supervise an agent. You do not edit English email.
We believe an AI can do almost all of this job. So your work comes down to one thing: teaching.
Judge the call, not the English
Every card carries a Chinese summary and its reasoning: the question, the source cited, what is missing.
One line in Chinese is one lesson
Say one line in Chinese; it rewrites the customer-language reply and keeps the correction as a rule.
Teach until it stops asking, then let go
You delegate by category. Refunds, compensation, complaints and chargebacks always stop for a person.
Demo animation · invented data
You
Core capabilities
Every channel, one card queue
Email, live chat and Amazon buyer messages each go out under their own platform rules and arrive in one card queue.
Learns from your mail and your site
Tone and limits from past email, plus FAQ, product and policy pages, become one knowledge base with every source shown.
Policies pass a trust audit first
Policy pages contradict each other or go stale, so they must clear a trust audit before auto-reply may cite them.
Supervise in Chinese, reply in the customer language
Type or hold to talk, in Chinese. You manage a support agent, not an English email editor.
From rehearsal to auto-reply
Auto-reply is open on every plan. The gate is knowledge maturity and safety rules — refunds and complaints need a person.
Live chat on your storefront
The same agent answers your storefront chat, says so honestly when it needs a human, and follows up by email.
Everywhere you work
One queue, and it counts wherever you handle it
The cards that need your call come to where you already are. Web and phone are one queue — teach on either and it learns the same.
Web workspace
The full card queue, knowledge base and delegation settings, for the complicated cases.
Card deck on your phone
Swipe right to approve, left to reject, up to snooze, hold to talk in Chinese.
Desktop companion
A small robot works quietly in the corner and raises a sign only when it needs you.
No per-seat fees
Anyone can teach the agent. You never buy another seat to add another pair of eyes.
In-group approval cards, available as an add-on
LarkDingTalkWeCom
Right approves · left rejects · up snoozes
Newest capability · Amazon channel
Amazon buyer messages, spoken by the platform's rules
A buyer message is not ordinary email: links get stripped, emoji count as a violation, and silence past 24 hours is a miss. Those rules live inside the send gate.
All 22 marketplaces recognised on the way in
Inbound knows the messaging domains of all 22 marketplaces, so buyer mail is identified and queued on arrival.
A hard outbound gate: non-compliant replies never send
External links, emoji, oversized attachments, template-speak — fail any one and the reply does not leave.
A 24-hour SLA clock that escalates in three steps
Every message carries a 24-hour countdown that escalates in three steps; response rate comes from closed cycles.
Outbound gate checks
Link blocking
Any off-platform link in the body is held back — shorteners and hot-linked images included.
Attachment validation
Type and size are checked against platform limits, so an oversized file is stopped before the send.
Template-speak detection
Wholesale copy-paste boilerplate gets flagged, so the reply still reads like a person wrote it.
SLA countdown
Once the clock reaches its final stage, the reply is pushed to the front of the queue.
The response-rate card is the same component the app renders; the figures here are demo data.
Not another Zendesk
They hand you a system to configure. We hand you an employee that learns.
Ticketing tools move your process into a console, and the deeper you configure, the more you live in their database. We took the other road: the agent reads, learns and works on its own, and your mail stays in your mailbox.
Ticketing tools like Zendesk / Freshdesk
KefuAgents
Getting started
Ticketing tools like Zendesk / Freshdesk ·Weeks of implementation: forms, workflows, triggers, macros, permissions — all before anyone starts.
KefuAgents ·Paste one store link: brand, catalogue and policy pages read, past mail imported, live the same day.
Where knowledge comes from
Ticketing tools like Zendesk / Freshdesk ·You write the help centre and maintain the macros — it knows exactly as much as you wrote down.
KefuAgents ·Past email and site pages are ingested automatically; the agent proposes, you confirm.
Who does the daily work
Ticketing tools like Zendesk / Freshdesk ·Agents read and write every ticket by hand, with AI suggesting from the sidelines.
KefuAgents ·The AI runs the loop — triage, order lookup, drafting, sending. You say yes or no on a card.
Management language
Ticketing tools like Zendesk / Freshdesk ·English dashboard, editing English mail line by line.
KefuAgents ·Chinese summaries and Chinese instructions; it writes the reply in the customer language.
Who holds the data
Ticketing tools like Zendesk / Freshdesk ·Conversations and knowledge live in the platform database; switching tools means starting over.
KefuAgents ·Support mail stays in your own mailbox, behind exactly one KefuAgents folder.
Going live
Ticketing tools like Zendesk / Freshdesk ·Finish the setup and it starts sending for real; mistakes surface afterwards.
KefuAgents ·Shadow mode rehearses first. You delegate category by category after seeing real output.
Pricing
Ticketing tools like Zendesk / Freshdesk ·Per-seat fees, then per-AI-resolution fees on top.
KefuAgents ·No seats: priced by brand AI employees, email scale and AI usage.
Launch workflow
Four steps, and the first three are free
1
Paste your store
One input box. Submitting it starts reading your brand, catalogue, policy boundaries and help centre in the background.
2
Connect the mailbox
Type the support address; server settings come from your domain's MX records, so you only add a password.
3
Watch it rehearse
After learning your email history the AI starts drafting without sending. You judge it against real customer mail.
4
Delegate category by category
You approve which kinds of mail it may send on its own. The risky kinds always stay with a person.
Shopify / Amazon order contextWebsite FAQ and policy learningStorefront live chatMCP · bring your own AI agent
Pricing
Finish the loop for free, then decide whether to pay
The free tier includes a one-time grant of AI credits — enough to run the whole setup loop at no cost: paste your store, get it analysed, connect the mailbox, learn your email history, and watch the AI rehearse against real mail. Pick a paid plan once you have seen it answer well.
Auto-reply on every plan, the free one includedWeb workspace, phone card deck and desktop companion on every plan — no per-surface upchargeNo per-seat pricing; team member counts come with the plan
Free trial
$0
Run the whole setup loop without paying: paste your site, connect a mailbox, learn from past email, and watch the AI work in shadow mode.
Indicative scale: 100 emails/mo · 1 brand AI · 1 inboxes
A one-time 300 AI credits at signup — yours to keep, they never expire
- 300 credits at signup — no expiry, no card
- Full-site analysis and knowledge bootstrap
- Learns from your last 100 emails
- Shadow mode first, so you see the quality
Starter
$15/mo
For new DTC sellers — auto-reply is included from the first tier.
Indicative scale: 600 emails/mo · 1 brand AI · 2 inboxes
450 AI credits every month, valid within that billing cycle; topped-up credits never expire
- AI email takeover
- Supervise the AI in Chinese
- Auto-reply
- Learns from your last 1,000 emails
Professional
$45/mo
For stores with steady support volume — live chat and knowledge conflict checks included.
Indicative scale: 2000 emails/mo · 1 brand AI · 4 inboxes
1350 AI credits every month, valid within that billing cycle; topped-up credits never expire
- Live chat, standard (1,000 sessions/mo)
- Knowledge conflict detection
- Daily AI support report
- Learns from your last 5,000 emails
Flagship
$99/mo
For high-volume sellers that need stronger automation, auditing and team permissions.
Indicative scale: 6000 emails/mo · 1 brand AI · 10 inboxes
3000 AI credits every month, valid within that billing cycle; topped-up credits never expire
- Live chat, pro (5,000 sessions/mo)
- Advanced auto-reply policies
- Auto-reply audit trail
- Team permission management
Custom
Contact sales
For high email volume, multi-brand groups, strict privacy, or dedicated deployment — each brand runs its own account and back office, billed together.
Custom email scale · custom brand AI · custom inboxes
AI credits configured per contract
- Isolated database
- Dedicated deployment
- Custom ERP/WMS/logistics integrations
- Hands-on onboarding service
Indicative scale is a sense of magnitude to help you pick a plan — not a quota commitment. What you actually consume is settled on AI usage.
AI credits granted monthly with a plan are valid within that billing cycle, counted from your own subscription date; any balance you top up yourself never expires.
MCP access
Want your own AI to drive? Here are 17 tools.
MCP access exposes a REST API and 17 standard tools over MCP (2026-07-28 revision). Claude Code, Codex or your self-hosted agent can pull pending cards, search knowledge, submit drafts, approve replies and register webhooks — so support plugs into automation you already run instead of becoming another dashboard.
Read the developer docsFAQ
Make the core loop safe first, then delegate
The launch priority is inbox takeover, knowledge bootstrap, policy audit and delegated auto-reply. You approve the scope of automation category by category — never all at once.
What do I need to prepare before starting?
A store address and your support mailbox password. Product lists, categories, policy terms and mail server settings are all derived from your own site and domain; you only confirm whether they are right. We will never ask you to fill in a product sheet or list SKUs.
Will the AI mess with my mailbox?
No. It creates or reuses one KefuAgents folder and moves mail it judges to be support mail into it. Supplier mail, platform notifications and personal mail stay where they are — not moved, not marked, not answered.
I am not comfortable letting an AI send mail on day one.
By default it does not. Every new connection starts in shadow mode: it drafts and never sends. Once you have seen enough of its real output you delegate category by category, and refunds, replacements, complaints and chargebacks always stay with a human.
How far does the free tier go?
All the way through setup at no cost: paste your store for automatic analysis, connect the mailbox, learn your email history, and run shadow mode to judge real reply quality. Auto-reply itself is open on the free tier too — limited by knowledge maturity and safety rules, not by price.
My English is weak. Can I still run overseas support?
Yes. Incoming mail comes with a Chinese summary and every draft with a Chinese review copy. You instruct in Chinese by typing or holding to talk, and the AI writes the final reply in the customer's language. You never edit English word by word.
How many brands can one account manage?
An account can manage several brands, but each brand is its own AI employee with its own knowledge base, inbox takeover and delegation policy. Mixing brand tone, return policies and automation limits is the single biggest source of wrong-reply risk.
Paste your store and see how much it reads
Reading takes under a minute. When it finishes you can see your own brand, catalogue and policy boundaries laid out — and this step costs nothing.