It answers the room, wherever the room is. From what you actually know. You keep the last word.
One copilot across a live show, a stream, a subreddit and your follow-up inbox. It answers from your listings, names the fact it used, and hands you the reply.
For the one person running it all · read-only on sessions you do not own · a human approves every reply, at every rung
2.00s
is the budget a drafted reply is measured against. We miss it.
It no longer reproduces on this build, so we do not quote it. Read your own: every breach is counted, never averaged away.
$0.06
a minute, measured once. Not a price.
One agent per session, metered by the minute, with the meter beside the queue. $0.52 for nine minutes.
6
deterministic guards. A reply passes all six, or it is held.
Rules over the catalog as it stands, not a second model asked to be careful. Our own accuracy numbers proved little and are not quoted.
Where it answers
The same copilot, in four places that behave nothing alike.
A live show runs at the speed of the block. A Tuesday thread still deserves a real answer on Friday. Both speeds, the same guards.
Live commerce
- eBay Live
- Whatnot
- TikTok Live
Answers from your listings, then proposes what follows: a markdown, a stock fix, a listing ended.
What it will not do
Only eBay Live acts on listings: your own show, after preflight, 90-second undo. The others draft only.
Creator streams
- Twitch
Twitch chat on a channel you connect. A stream's chat is a room like any other.
What it will not do
Key-gated, not connected in production. Clip, poll and pinned message are written and tested but not wired in.
Communities
Watches the subreddits you choose and answers with the thread above the question in view.
What it will not do
Draft-only in the code, four times over. You post it, under your own name.
Your follow-up inbox
- Follow-ups
Everyone who asked and did not buy. One reply each, re-checked against the catalog now.
What it will not do
A blocked follow-up is never written down, so everything waiting has already passed the chain.
Two tempos, one set of rules
Live gets a bounded session: attach, answer, end, report. Async gets a standing watch and an open queue.
What is running today
eBay Live is what we run; Whatnot reads the same way. The rest need their own credentials.
Who sends it
You are the sender. Nothing here posts for you: the copilot reads, drafts, cites and checks, and a human puts the words in the room. An approved reply is recorded, audited and handed back to you to copy. eBay Live publishes no chat-post API and the scraped surfaces have no send path by construction; Twitch's is written and tested against its API and is not wired into the app that runs. Rooms you do not own carry a posting switch that starts off, stays off until you turn it on for that room by name, and cannot be turned on at all where the surface is draft-only — and nothing in this build acts on it yet. That is a boundary we chose, not a feature we owe you.
Who it is for
One person, three jobs, four rooms.
A brand's live show has a producer, a moderator and a merchandiser. You are all three, on camera — and the questions outlast the show.
Questions scroll past
Every one was a buyer who was close.
Edits land late
The markdown, the sold-out lot, the wrong pin.
Fast answers go wrong
A stale price is a refund tomorrow.
Then it all disappears
The people who asked and left, and what your listings should have said.
Before · during · after
Three phases, everywhere. What fills them depends on the tempo.
Prepared, worked, reported on — or, where nothing ends, a watch and a queue.
01 — before
It prepares, and tells you what it could not.
No session should meet its first question with an empty catalog. It prepares one, and grades itself item by item.
- Resolves the seller behind the show and builds a catalog of up to eighty listings.
- Comparables are asking prices, labelled as such — the sold-price feed is limited-release and this keyset lacks it.
- Gives the show its own agent, with the lineup in its knowledge base and the never-say rules armed.
- Nine named readiness checks, plus a fit check for a catalog belonging to another auction.
without a session
Nothing to prepare. You name the subreddits, and posting stays off — on Reddit, permanently.
02 — during
It reads the same page your buyers read.
No private API, no special access. It opens the public player, reads the chat and lot card, and hears you if you share the tab. Lots are versioned as the auction moves.
perceive
Buyer chat, the lot card as the block moves, a camera frame, and your voice. Audio needs one click.
ground
Exact lookup on a resolved lot, then keyword and trigram search. Nothing resolved and it abstains.
check
Six guards against state re-read at check time. A guard that throws fails closed. Editing re-runs the chain.
propose
One keystroke clears a reply and audits it; you put it in the room. Listing changes need a preflight and stay reversible for ninety seconds.
without a session
No console, no clock, nothing to end. A standing watch, an always-open queue, and a copy button where the console has send.
03 — after
It reports, and then it owes people answers.
Five sections and a to-do list. The counts are measured; the conclusion is prose from the agent, so counts come first.
- Comments, questions, answered rate, latency, cache hits, what each guard blocked, and whether the audit verifies.
- Every unanswered question, with how often it was asked, and a button that answers it into the catalog.
- The session plays back: audio, transcript with emotion and intent, and the frames the agent read. Deleted with the session.
- Everyone who asked and did not buy leaves with a drafted answer, re-checked against the catalog as it stands now.
without a session
A watch never ends, so it has no report. What it leaves behind is the record in Drafts.
What one report looks like — the shape of the numbers, not a measurement of a particular show.
77%
grounded from the catalog — what the copilot could answer
23/31
sold lots that had a question answered
11
gaps, each with an Answer button
One number it will never contain
Wrong replies that reached a buyer.
A reply this system judged correct is the one it cannot mark wrong. So you flag them, we count what you flagged, and the report says plainly that the number is a floor.
The report shows both figures for your own session, side by side, and says that neither is the whole truth.
The session that made the inbox exist
One real fragrance auction — session ebay_47tK1SX0VsiHEXN1: 190 comments, 60 answerable drafts, 29 buyers — none sent. The inbox makes one draft each, re-run through the same guards.
What makes it different
It listens and looks. Most copilots only read.
A chat bot sees the chat. This one hears the host, sees the lot, and knows which listing version every fact came from.
It hears you
Share the tab's audio and your voice arrives as distributions — never one label pretending to be a fact.
It sees the lot
A camera frame, read by the same agent. Shown to you, and never counted as a citation.
Stale is provable
Every fact carries the listing version it was read at. A markdown makes old replies unreachable.
One chain, every room
Same guards on every surface. The console changes shape between them; what may be said does not.
Guardrails
It stops the answer that would be wrong.
No model is ever asked whether a reply is safe. Six deterministic guards run on every draft, against the listing as it stands now.
the seller marked it down 1.4 seconds ago — the listing is now v2
Mia_k, the Chicago Reimagined AJ1 size 10 is $412.00.
Blocked by the price guard
expected $370.00 (v2) · found $412.00 (v1)
Two and a half seconds. The lot was marked down and the draft quoted the old price; the guard compared listing versions and stopped it. Run it yourself: npm run demo:stale-price. Read past the block — the script repairs and sends, which the shipped pipeline does not do.
Every amount is a fact at the listing's current version.
No "last one" unless a cited lot has one left.
No delivery promises, off-platform payment or blanket authenticity.
A citation that was never retrieved is worse than none.
Your voice guide: short, no markdown, no hype.
No contact details into public chat.
What they do not do
- Judge whether a reply is good.
- Verify entailment — grounding checks connection, not proof.
- Catch what retrieval never surfaced. Abstention does that.
- Police what you say on air.
- Enforce a subreddit's own rules. We read them and count them; the guard that would judge a draft is built, not fed.
Where they run
Fifteen of the seventeen never-say rules are armed on the agent itself, so they hold on voice and the embed too.
Autonomy
It starts as a copilot and earns every rung after that.
Each level unlocks on evidence from your own shows. A rung changes how much clears without you asking, never who sends it.
Observe
Classifies chat. Suggests nothing.
Suggest · you are here
Drafts everything. You send it.
One-tap
Pre-approved for a single keystroke, after three clean shows.
Auto-clear
Allow-listed topics clear themselves once every guard passes. Never price, never discount.
Auto-act · locked
Stock fixes and markdowns with undo. Locked until rollback rate holds across five shows.
What it costs
One agent, metered by the minute. We show you the meter.
Nine minutes of a watch auction ran at $0.06 a minute — roughly $7 over two hours. An observation, not a rate.
$0.06
a minute, on the one session we measured
≈ $7
two hours, if that holds
Point it at tonight's show, or at the thread from last Tuesday.
One account, one box. A session you do not own is read-only: it drafts and reports, and never writes to a listing.