0.1.27 2026-10-11
From 0.1.20 on, changes a customer or operator can see are listed on this page. Earlier releases are not backfilled here.
0.1.27
The homepage title is centered. The product window shows the visitor conversation on the left and the sources behind the answer on the right. A narrow screen keeps only the conversation. Live numbers are four cards. Corners are rounder, buttons and cards are softer, and the grid on the page background is gone.
The website, Mac, Windows, iOS, and Android icons are generated from one headset-in-a-bubble drawing. The dark fringe on the rounded corners is gone. The green and the cream background stay as they were.
Starting the Mac agent desk no longer raises the system login-keychain prompt while it reads the saved session. If the keychain refuses, the session stays in an encrypted file on the machine, not as plain text. Windows and Linux still use the system credential store.
Choose the kind of offer, then the plan and the discount, then the dates, then the words a customer sees. A monthly gift lists only monthly plans. A yearly gift lists only yearly plans. The name and the badge can be generated from the offer. Inside the platform console, the page keeps a single scrollbar.
Platform gates and status maps are now the onboarding workbench. Each platform shows progress, what is stuck, and the next action. A platform that is not scheduled yet appears to companies as coming soon. A company can ask for a pilot. With one governor, confirming a gate is a single step. With two or more, the person who submitted it cannot approve their own request.
Apple and the Android stores start with the application id already taken from the project. WeChat Mini Program setup is four stages: project check, preview upload, a hands-on trial, then a manual review. The preview QR code can be opened from the campaign. Submitting for review and publishing still happen in WeChat.
0.1.26
The top of the page is Mac and Windows, side by side. When a build exists, the button is a download. macOS is currently a notarized native universal installer, Windows a zip, and Android an APK. Candidate builds show their signature status; check the SHA-256 before installing. Once an Electron release is published, the download button prefers it.
iPhone and iPad go to the App Store or TestFlight. This page does not offer an installer, and it does not show an Apple team id. The browser agent desk is still there if you do not install a client.
Apple, Huawei, Honor, Xiaomi, OPPO, vivo, Tencent App Store, Google Play, and WeChat Mini Program each have their own fields, with a note on where to get them. An incomplete setup can be saved, but it cannot be submitted, and the page names what is missing. Secrets are stored encrypted; the page shows only a fingerprint. A connection test has to pass before the account is ready.
A live submit checks the installer checksum, then uploads it and sends it for review. The app needs to exist in that store first. A WeChat Mini Program is uploaded as a developer build; review and release still happen in WeChat. If live submit is off, or a channel is still incomplete, the button stays disabled and says why.
Stores that report review status are synced: all live, something failed or was rejected, or still in review. WeChat and Google Play have no status API, so a channel can be marked live or rejected by hand. If a submit is interrupted, it is not uploaded again automatically, so the store does not receive a second copy. A failed channel can be retried on its own. A channel that already went through is not submitted again.
0.1.25
On PHP 8.1, a non-streaming model call no longer fails the whole turn. When a trial run did produce an answer and the facts did not check out, the step says the draft was replaced with the original answer. It no longer looks like the model call itself failed.
Answer checks and trial runs no longer create customers, leads, conversation records, or visitor memory. They still do not open a ticket or add a charge. Older evaluation rows already in the customer list are not deleted automatically. A cleanup can be previewed first.
Asking what KeBao costs answers with the current plans: Basic is 100 yuan a month, Pro is 299, Enterprise is 1000. Asking why a free trial still charges money explains the trial credit, instead of how to buy a token pack. Saving a question in the console keeps an already approved public wording.
The contact picker shows a name, or “visitor” plus the short code when there is no name. You can search by name, company, or that short code. The customer list finds the short code shown on screen. A global search no longer matches someone just because their internal id happens to contain the same characters.
A failed distribution can be retried. A campaign can be archived, or copied into a revision and sent again. An iOS ipa and an Android aab can be the real build. A channel can be marked optional, so a failed precheck warns instead of blocking the required channels. The detail view shows the timeline and approvals. Leaving an account secret blank keeps the saved value.
“First month free” appears only while a trial is actually running. Otherwise the page says no plan is active. A health check opened from the admin host no longer fails because that host is not a connected site. A platform account opening company profile no longer sees a save button.
0.1.24
A site can override the platform duty policy: whether humans are on, ratings, assignment, queue timeout, concurrent chats, and ticket time limits. Leave a field empty to follow the platform, or clear every override at once. How an agent is judged offline stays with the platform.
A company admin can limit which regions members may sign in from. Admins themselves are not limited. The company can also set a default interface language, or leave it to follow the platform.
Platform billing is now plan pricing, add-ons and usage packs, tier quotas, campaigns, and payment setup. The old pricing address still works. From a plan card you can open the matching campaign with the form already filled.
Fields, tags, pools, teams, phrases, and the recycle bin appear according to the permission each one needs. What you can see, you can save. A company admin can manage teams. A site with only one member is told that teams are not needed yet. The event queue in the recycle bin appears only when outbound webhooks are on.
Company admins and a finance role can open receivables, invoices, project costs, and commissions, and only for their own company. An invoice is picked from a contract, which shows how much is still billable. Accruals show the person and the project, and can be approved or paid in a batch. You cannot approve your own accrual.
Sign and template review can open an application, send it back, and filter by company or keyword. An existing cloud sign can be assigned from a list. Domestic signs accept only a company, a trademark, or a public body, and need an approved qualification. Resubmitting after a rejection updates the original application. A sign that already passed cannot be deleted outright.
A metric with no samples shows a dash, not zero. Tickets handed to a person and never picked up are listed on their own. Conversation records can be filtered by grounded answers. Answer evaluation runs the same reply path a visitor gets, without creating a ticket or billing. A trial run shows which sources were used, and can be saved as an evaluation case. If visitor likes are off, the improvement page says so and can turn them on.
0.1.23
Each site can publish one page with a short intro, services and prices, questions, a booking link, and contact details. Chat sits on the page, and the ticket source is marked “contact page”. Six templates are included, or one sentence can draft the page before you publish. The published page gives a link and a QR code.
Phone, email, WeChat, WeCom, WhatsApp, Telegram, LINE, and a URL are shared by the widget’s “more ways to reach us” and the contact page, up to eight. The page can use your own domain; the certificate and reverse proxy are configured separately.
The default is still AI first. A site can switch so that, during working hours with an agent online, the visitor goes straight to the human queue. With nobody online, or after hours, AI still answers. If the queue times out into a message, AI takes it back.
After a handoff, the agent sees a summary and a draft reply already in the box. They can use it, discard it, or regenerate it, or hand the thread back to AI without closing the ticket. A clear “I want a person” in Japanese, Korean, Spanish, and other languages goes straight to a person. The widget adds those languages, and Arabic runs right to left.
A site can connect its own mailbox. Unread mail is fetched each minute, and the same sender stays in one thread. AI and agent replies go back by email, so the visitor continues by replying. Bounces, auto-replies, and mail the mailbox sends to itself stay out.
A handoff can also email agents who turned notifications on, or push to Telegram. After an agent connects once from the duty desk, replying to that Telegram notice sends the reply back on the visitor’s original channel. The same thread is pushed at most once a minute.
A company requests a sign and a template in customer management, uploads proof, and waits for platform review before it is submitted to the SMS provider. The review result is written back. A sign already used or under review by another company cannot be used to send. The platform default sign stays available to every site.
The customer list uses one row for whose customers, and another for follow-up state. Both can be on at once. “Mine” and “overdue” show the same count as the list you open. A follow-up written by a person clears the due date; a note written by the system or AI does not. The recycle bin can restore a customer in place.
Deal cards lead with a phone number and whether follow-up is late. A deal can move to the next stage, or to any stage, without dragging. An unassigned deal can be claimed. The longest overdue follow-up sits at the top. While you look only at open deals, won and lost columns shrink so the board fits.
My tasks can change priority, owner, due date, and the description. Checklist items, dependencies, and watchers can be removed. The list can mark a task done, and more than 50 tasks page. Reassignment notifies the new owner.
The site overview shows five cards: contact entry, knowledge, model, handoff, and channels. Each card opens the place to finish it. Every plan can bring its own model, including after a plan expires, and that usage stays on the customer’s model account. The connection can be tested before saving. The privacy policy states that knowledge and visitor conversations are not used to train models.
0.1.22
Basic, Pro, and Enterprise can each name a chat model, fallback models, and whether to rerank. If the main model times out or errors, the fallback chain is tried. One model can hold several keys; a rate-limited key cools down. A tier without its own setting keeps the global default.
Enterprise can bring its own model endpoint. That usage is billed on the customer’s model account, not against the KeBao token balance. When billing is off, behavior stays as before.
With tool calls on, the model can search the knowledge base again with a new query, or suggest a handoff. If the visitor names a callback time and already left contact details, a follow-up can be saved. Retrieved passages can be reranked; a timeout or a missing rerank model keeps the original order.
When a site turns on image understanding and a vision model is configured, a channel image is described first and then answered from knowledge. If vision fails or is off, the conversation still goes to a person, and the image stays in the thread.
Visitor messages and model replies can be checked against blocked words, patterns, and an optional third-party reviewer. If that reviewer times out or errors, the message is allowed through so support is not stuck. A blocked outbound reply is replaced with a handoff line.
The widget can stop a reply that is still generating. Text already written is kept and marked interrupted. Several channel messages sent within a few seconds can be merged into one answer.
Feishu, DingTalk, Telegram, and Teams groups answer only when the bot is mentioned, and each person keeps a separate thread. A site can switch to a wake prefix, or share one thread for the whole group. Direct messages are unchanged.
An app already built on Dify, Coze, or Alibaba Model Studio can be an answer pipeline. If that app times out or fails, KeBao answers from its own knowledge, or hands the chat to a person. The model fee stays on the customer’s own platform account.
Each channel on each site records the latest inbound, outbound, success, and error. An alert fires after enough failures in one day. Store connections, status maps, and platform gates are managed in the platform console. A gate decides whether a tenant can turn that platform on.
Customer management adds follow-up activity, a recycle bin, and statistics. Desktop, mobile web, iOS, and Android read one menu, trimmed by role. The public site has its own privacy policy page covering what personal data is collected, why, and what rights apply.
0.1.21
Basic, Pro, and Enterprise. The first month is free and uses Pro features. When the monthly token allowance runs out, a token pack can cover the rest. Packs are not folded into the monthly allowance, they are not cleared at month end, and they expire on their own date. SMS is counted by message, email by message, and voice by minute. Included amounts are used first, then usage packs.
An upgrade credits the unused days of the current plan. A plan cannot be downgraded while it is still active. A capability that is not included asks for an upgrade instead of opening silently.
The pricing page picks the lowest tier that covers the seats, features, and tokens you select, and shows monthly and yearly totals separately. The prices follow the plans currently on sale in the admin console.
The homepage shows total conversations, tokens, connected sites, and visitors online now. The figures are totals across sites. They leave out evaluation traffic, site names, and conversation text. Quality figures stay hidden when there are not enough samples.
Duty-desk tickets use kinds such as complaint, incident, and service request. A status can move only to an allowed next step.
Apps, keys, and call logs for the Open API and MCP live in the company console. They no longer require a jump to system settings.
0.1.20
FAQ hits, documents, and vector passages become numbered references. One reply states which ones it used, how sure it is, and whether a person should take over. Prices, links, model numbers, and other figures that are not in those references are rewritten once. If they still do not match, the reply says a colleague needs to confirm it. The raw source text is not dumped on the visitor.
System settings can store answer pipelines. A site can use its own knowledge, or add knowledge from other sites. Sites without a pipeline keep the previous global setting.
Short closings such as “thanks” or “got it” are answered directly. They do not search, call the model, or attach links. Greetings do not get stray links either. If the visitor is upset, or the answer is unsure, handoff to a person stays available.
Company members can sign in by scanning a WeChat code. Platform admin accounts cannot. They still use a username, password, and one-time code.
After sign-up, a checklist covers the company name, verification, welcome message and persona, knowledge upload, embed code, a sandbox reply, and inviting the first agent. Colleagues join from an invite link.
Until the company is verified, SMS, email, outbound calls, full phone numbers, and exports stay blocked. Companies that already existed get 30 days.
Super admins and self-serve company accounts are separate. Platform sessions are shorter, and writes ask for a recent sign-in again. Super admins must turn on two-step verification.
Tasks, workflows, and operating figures are counted only after the site is selected. If orders, projects, or renewals are not connected, the page says they are unavailable. It does not show 0 as if the data were there.
Android does not start push SDKs that read device identifiers until the privacy policy is accepted. The desktop app only grants shell access to pages inside the installed package, and the duty desk stays on the KeBao production host.
0.1.19
