BrainCue

Vision — from interview copilot to ambient conversational companion

Status: direction set 2026-07-21. This is the v2 north star. The product spec is 01-PRD.md; the delivery plan is 10-ROADMAP.md.

1. The shift in one paragraph

BrainCue v1 is an interview copilot: it listens to an interview, detects questions, and streams grounded answer cues into a capture-invisible overlay. v2 keeps every one of those muscles but changes what the product is: an ambient AI companion for live conversations and activities. It hears what is happening on your machine (with consent), decides when it can contribute, and delivers help through whichever surface fits — silent overlay cues or its own voice. Interviews become one mode among many: candidate copilot, interviewer assist, meeting copilot, tutor, or an ambient companion while you work or game.

2. Why “ambient companion”, not a voice-mode clone

Generic voice chat is ChatGPT/Gemini home turf — we cannot out-latency or out-price them, so we do not compete there. BrainCue’s structural advantages are things a cloud chat app cannot do:

Advantage Why a cloud assistant can’t follow
Present in real conversations — system-loopback capture puts it inside your actual meetings, calls, and games They live in their own app; they can’t hear your call
Invisibly present — the overlay is excluded from screen capture (WDA_EXCLUDEFROMCAPTURE) Requires OS-level window affinity, i.e. a desktop app
Grounded in your corpus, locally — documents, notes, (later) memory stay on the machine; only retrieved snippets leave Their grounding is cloud-stored by design
User-owned intelligence — BYO key, no accounts, full local deletion Their business model is the account

The unifying frame: BrainCue is in the room with you. Voice chat is one of its surfaces, not the product.

3. What the user picks: an activity

Revised 2026-07-28. This section used to be a mode catalog, and the modes were shown to the user in a picker. They no longer are — see 18-ACTIVITIES.md for why that picker was deleted.

A mode is a preset over one shared engine (§4) — never a forked pipeline — and it is now an internal concept. The user says what the call is; the mode follows from it. That collapse happened because choosing a mode and choosing a Space kind were the same question asked twice, and they could disagree.

Activity You are… BrainCue… Runs Status
Meeting or call a participant quietly surfaces context, unanswered questions, action items, decisions meeting 🧪 shipped (Labs)
Project discussion talking about ongoing work recalls prior decisions and open threads meeting 🧪 shipped (Labs)
Interview the candidate detects questions, streams grounded answer cues framed as you interview ✅ shipped (v1)
Study or tutoring learning something surfaces the part of your material that bears on what was said meeting → tutor 🧪 partial
Personal dealing with your own life keeps the background straight, catches what you agreed to meeting 🧪 shipped (Labs)
Game · Just me playing / working / thinking ambient presence; speaks when it should, stays silent when it shouldn’t companion 🧪 shipped (Labs)
Something else anything else contributes only when confident meeting 🧪 shipped (Labs)

Practice (mock + sparring) is not an activity: it is preparation for an interview, in its own drill pages. It used to sit in the mode list, where picking it started nothing — it navigated away.

Two modes are still unbuilt: Interviewer Assist (you are the one asking — question suggestions, coverage tracking, a drafted evaluation) and Tutor (voice dialogue and drills; subject runs meeting until it lands). An activity whose mode is off is not offered rather than silently downgraded — starting a standup as an interview is worse than not starting it — so the roadmap lives on Home’s Labs strip instead of inside the picker you must answer to start.

The voice/summon layer (push-to-talk, spoken replies with barge-in), session summaries, and the review-first memory substrate are output/grounding surfaces over the one engine, not modes.

4. One engine, many modes (the core bet)

Every mode is a configuration of the same six-stage pipeline:

Sources → Transcription → Trigger policy → Grounding → Generation → Surfaces
 (mic,      (Realtime      (when should     (RAG over    (persona     (Cue Card,
 loopback,   GA STT)        I contribute?)   local docs   prompt,      voice, reports)
 screen,                                     + memory)    streaming)
 hotkey ask)

v1 already built one full vertical slice of this: realtime STT, a question classifier (which is simply a reactive trigger policy), the RAG retriever, streaming answers, and the overlay + TTS surfaces. v2’s foundational work is extracting the engine so a mode may only configure it, never bypass it — the rule that keeps six modes maintainable instead of six forked products.

5. Product principles

  1. Local-first, key-owner-first, provider-agnostic — data lives in SQLite on the user’s disk; AI calls use the user’s own key, which never reaches the renderer. OpenAI is the first provider, not the identity: the engine talks to a provider layer, and multi-provider support (Anthropic, Google, local models) is on the roadmap.
  2. Grounded, never inventing — contributions cite the user’s corpus; when there’s no match, say so and offer a safe framing instead of fabricating.
  3. Consent & transparency — listening starts explicitly, and the user can always see what was heard and exactly what left the machine.
  4. Silence is a feature — an ambient agent is judged by when it doesn’t speak; sensitivity is user-tunable and quiet is the default posture.
  5. One engine, many activities — modes are presets, not forks, and the user never picks one. They say what the call is; the mode is derived.
  6. Continuity is scoped, and memory belongs to the user — a conversation is kept in the Space it happened in, or not at all, so the tenth standup is grounded in the previous nine and one client’s history can never ground another’s call. Long-term memory about the person is a separate promise: off until consented, proposed rather than taken, and only ever recalled after the user approves it in the Memory section — visible, editable, deletable. See 16-CONTINUITY.md.
  7. Assist where allowed — unchanged v1 ethics posture: no anti-proctoring, no evasion, persistent “use only where AI assistance is permitted” reminder.

6. Brand

Keep BrainCue — “cue” generalizes perfectly past interviews (cue cards work for any conversation). Retire the “interview copilot” descriptor.

Element v1 v2
Name BrainCue Copilot BrainCue (appId/binary unchanged — no installer churn)
Descriptor “AI interview copilot” “ambient AI companion” / “the AI in the room with you”
Overlay “Cue Card” keep — it’s the brand’s best asset, now cueing any conversation

Open brand decisions (owner: maintainer): final tagline wording; whether the electron-builder productName changes now or at v2.0 release; README hero copy. The npm package name became braincue (2026-07-22) with the userData directory pinned to the legacy id so existing installs keep their data.

This site is open source. Improve this page.