The souffleur on your own machine. It hears the room, holds the entire meeting in one attention state, and sets the line you need one inch under your camera — before the question has finished landing. You keep looking exactly where you should be looking: at them.
FREE FOREVER. NO ACCOUNT, NO CLOUD, NO TELEMETRY. ZERO EGRESS BY DEFAULT — NOTHING LEAVES THIS MACHINE UNLESS YOU HAND IT A KEY AND FLIP THE SWITCH YOURSELF. THIS PAGE MAKES ZERO NETWORK REQUESTS, LIKE THE PRODUCT.
Forty minutes in, half your mind on the document you're presenting, someone says: "So — can you own the rollout by Friday?" You heard your name before you heard the question, and now the silence is yours to fill with whatever you can reconstruct while everyone watches you reconstruct it.
The question, mid-sentence: "…so, Bo, can you own the—"
The glass, three words in, before the sentence completes: ▸ AIMED AT YOU — rollout date · you flagged capacity risk at 00:12 · the number quoted in March was $40k
You, eyes on the camera the entire time: "Friday's tight — I flagged the capacity issue earlier. Here's what I can commit to."
This tool exists because of a stupider loop its author actually lived: stop the recorder so the file writes to disk, upload it to a transcriber, paste the transcript into a chatbot, re-explain the context — four or five times per long meeting, chasing a delta that was always forty minutes stale. The whole estate this software comes from began as the refusal of that loop.
You are the only person in the meeting who cannot also read the meeting.
Your attention is spent performing the conversation — that is what being in it means. Everyone else's notetaker reads the meeting afterward, when the moment is gone. And help through your ears would arrive on your slowest channel: speech runs ~150 words a minute, serial, while your eyes read 400+, skimmable, in any order. The glass is a bandwidth adapter — the meeting arrives at listening speed; your counsel arrives at reading speed, in a channel nobody else can see or hear.
PROMPTER is not something you ask. It has no chat box, no prompt bar, no send key — and for most of every meeting you will forget it is there. It is a resident: a small open-weight model on your own GPU that ingests both sides of the call word-by-word into one continuous attention state and judges, silently, at every completed thought: does the person at this desk need a line right now?
Both lanes — your mic, and the meeting's own audio captured per-application — word by word, as they land. Streaming captions included. Hour five remembers minute three.
The question aimed at you, caught mid-sentence. The number from earlier, with its timestamp. The contradiction with what was agreed. The answer sitting in your own brief.
Speaks, sends, joins, or acts. It has no send key by construction — the product that exists because send keys are the disease does not own one. The performance is yours.
Every word appends to a persistent attention state the moment it exists; nothing is ever re-read. A six-hour day fits resident with headroom — no sync, no refresh, no "processing…".
The boundary signal reads 0.97 at a completed thought and 0.02 mid-phrase, free, off the same pass that heard the word. The prompt fires where a colleague's whisper would: where a finished claim exists to answer.
A thin, click-through band placed just below the camera — so reading it is looking into the camera. And it is share-invisible by construction: screen shares and recorders cannot see it. Proven with a screenshot probe, not promised.
Lines the record supports render green with their citation. Anything the machine cannot stand behind renders visibly dim. A prompter that can feed you a wrong number with a straight face is a liability; this one refuses by rule.
Default-silent, an interruption budget you set, every hold logged with its margin. Most meetings the glass shows captions and nothing else — and that quiet is a decision on the record, not an absence.
The local resident is the whole point and it never goes away — it is the engine that hears everything and answers now. But like a hybrid drivetrain, a second motor can bolt on behind it: bring your own key, and the moments that deserve deeper thought get sent up — through a privacy gate, on your key, to the frontier model of your choice — and come back as a labeled refinement.
Sub-second, local, color-graded by its own measured confidence. You are never waiting on a network to be helped.
High stakes plus its own low confidence — that combination, judged locally, triggers the escalation. The cheap continuous mind spends the expensive discrete one.
Outbound text runs through an open-weight privacy-filter model running locally (OpenAI's released filter weights) plus a plain-regex floor you can read — names, numbers, identifiers scrubbed before anything leaves. On by default when a key is present; yours to toggle; text-only egress, ever — audio never leaves under any setting.
It renders on the glass with the ⟲ mark, its source, and its latency — "now that I've had a moment to think: here's the fuller answer." Human advisors have always worked this way. And if the deep tier refutes the reflex line, the glass says so plainly — the correction outranks the ego of the first draft.
A Cerebras-class key — thousands of tokens per second; the deep return lands in a breath.
A DeepSeek-class key — depth at pennies, arriving a few seconds later.
A Claude-class key — the strongest counsel available, worth its extra seconds.
Your key, your provider, your bill, your choice — any OpenAI-compatible endpoint works. And the doctrine holds either way: the spine never depends on the cloud. Remove the key and the full product remains. This is the estate's original two-tier design reborn — the fast tier used to be a cloud call too; now it's a resident, and only the optional depth rents.
Sometimes you don't want a whisper — you want to ask. Hold the hotkey: your mic to the meeting mutes, and your voice routes to the prompter instead. Ask it anything — "what did they quote us in March?", "summarize where we are", "give me the counter to that" — release, and you're back in the room.
The reflex answers on the glass immediately, color-graded like everything else — and if a key is present, the question escalates in parallel, so the deep refinement lands seconds behind the tentative answer. On a fast key, nearly together.
It's the question you'd have typed into your phone — without breaking eye contact, without the room watching you look things up.
The others hear a beat of silence, exactly as if you'd paused to think. Because you did — you just didn't have to do it alone.
Your name, your role, "what do you think" — flagged mid-utterance, before the question completes. In the live test it fired three words into the sentence. The silence never becomes yours.
"The number we quoted them" — from minute 12, at hour two, with the timestamp. Nothing is ever re-read; it was never put down.
What's being proposed collides with what was agreed forty minutes ago. The glass shows the agreed line, cited. You decide whether to say so.
"Actually — scratch that." The standing summary re-forms in a quarter second; the old plan dies mid-thought, the way it should.
It walks in already holding your prep notes, the attendees, last meeting's tape. Asked about something you wrote down Tuesday? The line is there.
Owners, dates, action items accrete silently into the one artifact worth having when it ends — and next meeting, it walks in briefed.
The escalated moment comes back refined — labeled with its source and its lateness, integrated where it belongs, and flagged loudly if it overturns the reflex line.
Every serious theater for four hundred years has had a box built into the front of the stage. Inside sits the souffleur — the prompter — who holds the entire script, follows every spoken line, and stays perfectly silent for the whole performance. Their one job is the moment an actor dries: the line, whispered, just loud enough for one person.
The audience has always known the box exists. They just never see inside it — and no one has ever called the actor's performance fake, because the performance is the actor's. That is the register this tool lives in: present the whole time, silent almost all of it, and yours.
And it records people — so the doctrine ships in the product, not the fine print: one-party consent is the legal floor; disclosure is the doctrine. Your jurisdiction's rules are yours to follow. The tape stays on your disk and belongs to you.
Meeting platforms now offer AI participants, and they are useful — for the meeting. Note the allegiance: the room's AI is the room's. It summarizes for everyone, records for the host, and gives you nothing the others don't see.
The platform's AI works for the meeting. PROMPTER works for the person at this desk. It is counsel, not a notary.
No host toggle, no platform feature, no permission architecture stands between you and your own machine's audio. If the host disables recording bots and AI participants, the room's AI vanishes. Yours doesn't — it was never theirs to configure.
Zoom, Teams, Meet, a phone on speaker, a conference room — anything this machine can hear, it can attend. One tool, every room, no integration.
This is the property the whole estate began from: your own local recording answered to you when nothing else did. PROMPTER is that property, grown a mind — and the disclosure doctrine above travels with it into every room.
Every meeting's transcript, decisions, and judgments on an append-only local record. Next meeting on the project, it walks in briefed.
A text file it prefills before the call: the project, the people, what you want out of the hour. The model arrives already knowing.
Every dismissed prompt and every used one is a labeled example — on your machine, in your corpus. The disposition re-tunes toward your sense of what deserved the glass. That record never leaves, and it is the one thing nobody can clone.
PROMPTER rides the FUSOR-1 resident substrate. These are bench ratings — measured, dated, receipted in the reference build's public runs ledger:
| Quantity | Rating | Receipt |
|---|---|---|
| Word spoken → caption on glass | ~350 ms | streaming partials, live |
| Question aimed at you → flag | mid-utterance | fired 3 words in |
| Thought-boundary detection | 0.97 vs 0.02 | ingest logits, free |
| Completed thought → judged | ~44 ms | one-token probe |
| Final word → advice on glass | 457 ms | live end-to-end |
| Reversal → summary re-formed | 232 ms | mid-decode abort |
| A day held resident, no fold | ≥ 98,304 tok | multi-hop recall |
| Share-invisibility | proven | screenshot probe |
REFERENCE BENCH: ONE RTX 4070 Ti SUPER (16 GB) · 9B-CLASS OPEN WEIGHTS · WINDOWS. HEADSET REQUIRED IN V1 — SPEAKERS AND ECHO-CANCEL ARE A V2 PROBLEM AND THIS PAGE SAYS SO. THE REAL-CALL HARDWARE GAUNTLET (BLUETOOTH MID-CALL, TEAMS AEC COEXISTENCE, DEVICE CHURN) IS THE OPEN GATE BEFORE V1 — RATINGS ABOVE ARE BENCH RECEIPTS, NOT FIELD CLAIMS. THE HYBRID TIER (§04) AND PUSH-TO-ASK (§05) ARE DESIGNED, NOT YET BUILT, AND WEAR THEIR AMBER TAGS UNTIL THE TAPE SAYS OTHERWISE.
PROMPTER is a fit: the open resident-kernel substrate pointed at the desk seam, with a disposition tuned for it. The loop, the bus, and the laws live at fusor1.com.
In a tool that hears your meetings, "trust us" is not an installer. PROMPTER is MIT-licensed source you build on the machine it will run on, with open weights you choose. No license server, no account, no vendor in the loop.
STATUS: REV 0 — THE RESIDENT CORE IS BUILT AND MEASURED IN THE REFERENCE REPO; THE PROMPTER FIT IS BEING EXTRACTED FOR RELEASE, AND THE V1 GATE IS THE REAL-CALL HARDWARE CHECKLIST. COMING SOON MEANS COMING — STAR THE REPO OR CHECK BACK. EVERY CLAIM ON THIS PAGE STAYS PINNED TO WHAT THE TAPE CAN BACK, AND THE AMBER TAGS COME OFF ONLY WITH RECEIPTS.
Turn-based assistants meet you after the moment. The prompter is the first tool built for during: a whole meeting held in one attention state, a whisper on glass, a deep tier when you want it — and nothing leaving your machine unless you send it.
REEL asked how memory survives. KEEL asked how programs share a core. BRAIN asked what closes the loop. CORTEX asked what remembers. HYPERCELL asked how many. FUSOR answered when — and PROMPTER is the machine the answer was first built for.