A digital human with a face, a voice and a job — and a full content pipeline behind it. Here is the whole thing, in four parts.
A face, a voice, a character and a job. Everything below it is the product it has access to.
A persona that can hold a live conversation — see it, hear it, talk to it — with the whole application behind it. It is not a chat window with a picture on top: while it talks it can read what you published, see what is scheduled, check which accounts are connected and start work.
Three routes. Vendor video is the default: photoreal, rendered on the provider's side, billed by the minute. A browser-rendered figure draws a 3D model on the viewer's own device and costs nothing per session — real, unfinished, and off until a platform admin enables it. Voice-only has no picture at all, which is what a phone call is.
Upload a VRM or glTF model you already own, or make one from a single photograph in the back office — no account with the model vendor required. A figure keeps the same face across sessions, so a returning visitor meets the same person.
Pick a voice from your asset library — cloned voices carry a consent record — then choose a character preset and move four dials: warmth, formality, initiative and humour. Answer length is set with them, because a companion and a support desk disagree about what a good answer is.
The job it is given decides how it opens, what it volunteers and what it must never say: a pre-sales assistant may not invent a price, an after-sales desk may not promise a refund it has not confirmed. Fourteen jobs are written, and social content operations is the one this phase stands behind; the other thirteen — customer service, pre-sales, after-sales, bookings and reception, course and community Q&A, product help, lead capture, guided tours, fan liaison, companionship and more — appear in the picker marked coming soon, and arrive in the next phase. Shipping a job before it is finished means somebody meets the gap in front of their own customer.
Your brand, products and policies, indexed into a knowledge space and retrieved by meaning at the moment it is asked. It answers from your material rather than from a paragraph you pasted into a prompt, and it says so when it does not know.
It can list your recent posts, tell you what is scheduled, report what is mid-flight, name which accounts are connected, and start a content task from what you just said. Anything that creates work can be set to ask first: it puts a proposal on a list and a person accepts or rejects it.
In the app, on your own website through an embed key, or by telephone. An embed key is public by design and can only start a conversation — its safety is the ceiling on it: minutes per session, sessions per day, how many at once, and which domains may frame it. Telephony needs a SIP trunk from a carrier and a LiveKit SIP service; until both exist the app says so plainly rather than failing quietly.
Live conversation is metered by the minute: a model call, a voice, the room, and — on the vendor-video route — the avatar stream, which is most of it. Credits are reserved before a session and settled against what it actually used. Choosing voice-only or the browser figure removes the largest line.
The same digital human, seen from the two sides of a transaction.
Browse what is on offer, see the headline, what it is for, its rate per minute and its rating, and hire it by the minute. A listing nobody has rated yet shows no rating rather than a zero — no score and a bad score are different facts.
Today every listing is ours: the seller side is built and runs, and it opens to other organizations with white label. When it does, a listing needs a headline, what it is good at, tags and a rate per minute — or no rate, and it charges the platform rate.
The cost of serving a minute comes off the top; what is left is the margin. The seller keeps 70% of that margin and the buyer gets 5% back as credits. 70% of margin is worth more to you than the same headline percentage of gross elsewhere, because the platform has already absorbed the cost of running the conversation — and it means a seller can work out their own earnings from their own price.
Earnings arrive as credits and can be spent immediately. Where the platform has cashing out enabled, only credits that were earned can be withdrawn — never bought or granted ones — above a minimum balance and through a verified payout account.
Every conversation is recorded as a session: who, how long, what it cost, and a transcript summary the owner can read. A buyer can rate what they hired, and that rating is what the next buyer sorts by.
The content pipeline. Talking to the digital human is one way to start it; the create page is the other.
Create → pick a platform preset (fixed format, one platform, fewest choices) or Freeform Creation (full control: any form, any platforms, batches and series). Write one topic direction — or tick "use my topic directly" to skip topic generation entirely. A preset also carries its own defaults for language, region and the topic sources it offers.
Each task locks its target locale at creation, so a Japanese task stays Japanese all the way to publishing even if you switch the UI language later. Script matters as much as language: Traditional Chinese is carried as Traditional, not as "Chinese". Region shapes references, units, currency and spelling, and seeds trend and holiday lookups.
Seven sources, chosen per task: write it yourself, give a URL to read, give keywords, pull from trending boards for your region, retrieve from your own knowledge base, repurpose something you already published, or take the next slot on your content calendar. Whichever you pick, the topic gate is where you decide what actually gets written.
Four stops, each in front of something expensive or irreversible: topics, the draft (before any media is generated), the storyboard for video (before minutes of rendering), and the finished post per platform (before publishing). Reject with notes and the AI regenerates under your constraints — the notes are the instruction, so "too corporate, and the second point is wrong" produces a fix rather than a different mediocrity. Gates you trust can be auto-approved per task; the choice is recorded on the task and every automatic approval is audited.
Image, comic, audio, music and video forms generate real media. Videos get an editable scene document: the editor covers scenes, narration, BGM and captions, and its collapsed panels add a multi-track timeline, layered overlays and precision trims. Recurring characters keep one face across episodes through a locked character sheet. Approving the finished post approves the media inside it — there is no separate media sign-off.
Upload from your computer, or scan the QR code and send photos and video straight from your phone. Each asset carries a role — fact source, inspiration, visual style, voice or music reference, brand asset, must-use — and that role decides how the pipeline reads it. Documents indexed into a knowledge space are retrieved by meaning at generation time, so brand facts reach the writing without being pasted into every brief.
From the finished-post gate or the publish calendar: post to a connected account, schedule it, keep a manual reminder, download the paste-ready text plus media files, email the package, push it to your webhooks, or notify your team chat. The calendar's + adds a package to any day and exact time, and shows what is already booked so a week does not accidentally ship three posts on Tuesday.
Where a platform offers no usable API, an approved account can hand the finished post to a cloud browser that posts it the way a person would, using a login vault and a pinned residential IP so the account looks like itself every time. It stays behind the same approval as everything else: nothing reaches a channel that a human has not said yes to.
Comments and messages from connected accounts arrive in one inbox. The digital human can draft the reply, or answer it itself under the job you gave it, and hand the conversation to a person the moment it stops being useful.
Compare performance by language, template, form and workspace, watch what competitors in your feed are doing, and roll a campaign up into a report. Per-post metrics collect from connected providers or manual backfill. Nulls stay nulls — we never show a fake zero, because "no data yet" and "zero views" are different facts and only one of them is a problem.
Who pays, who can see what, and how other software talks to it.
Subscriptions grant entitlements and monthly credits; each variable-cost operation shows its credit price up front, reserves before running and settles what it actually used. Failed operations release their reservation, and a job delivered twice is charged once. No plan sells unlimited generation, because the cost of a plan is fixed and the cost of a video is not.
An organization holds billing, credits and members; workspaces separate the actual work — one per brand, client or market — so content, assets and analytics never mix. Teams are an optional grouping that grants access to workspaces without owning anything. Roles run from owner and billing at the organization down to workspace admin, editor and viewer.
Everything the UI does, the public API does too: organization API keys with scopes, per-workspace restrictions and rate limits, plus outbound webhooks that push content, media and publish events to your own systems. Keys are stored only as hashes, can expire or be revoked, and every request is metered and auditable.
The models, the media engines, the storage and the publishing connections are adapters chosen per task, per user, per workspace or per organization, in that order. Which provider ran a job is recorded on the job, so a change of provider is visible rather than mysterious.