NEW — Real-time fact check · Meeting memory AI Meeting Copilot Local-First Zero Setup macOS (Apple Silicon)

Stay focused on the meeting. Bob handles the record
and context in real time.

Real-time fact check corrects wrong numbers on the spot, and meeting memory resurfaces past decisions.
Live 2-pass transcription × a built-in local AI that knows your notes — install and you're done.

NEW 01 · Intelligence

It checks facts as you speak,
and remembers past meetings

"Did we say it grew 20% last year?" — Bob detects claims with numbers, years, and product names, verifies them instantly, and resurfaces past decisions with sources the moment a related topic comes up.

Real-time fact check

A local gate picks out claims and AI verifies them — verified, incorrect, or uncertain, with evidence on a card. If it's wrong, the correct fact comes too.

Meeting memory

When a meeting ends, decisions, facts, and plans are learned into a knowledge graph. Next meeting, related topics bring back past conclusions with a "Last meeting · date" source.

Per-meeting toggles

Knowledge-base suggestions · fact check · meeting memory — switches on the start screen. Fact check is off by default, and even when on, only gated utterances are minimally verified.

Meeting memory and note suggestions run fully on-device. Only fact check is opt-in — with your own API key, only when needed.

Meeting intelligence

0
sec

From speech to first-pass text —
perceived latency of 2-pass streaming

Final transcription accuracy
after the offline refinement pass

0
cloud uploads

Audio, transcripts, notes — all on your Mac.
Zero data leaves the device

0
gate layers

Prefetch → heuristic gate → LLM synthesis.
Only what matters gets through

02 · Live

While the meeting flows,
this is what the screen does

A faithful replica of the app's Live session screen. When an utterance is finalized, Bob searches the knowledge base — only gated suggestions and fact-check cards appear in the intelligence panel on the right.

Meeting intelligence

03 · Translate

Speak any language —
read it in yours, instantly

Pick a conversation language when the meeting starts, and the built-in local AI translates other languages in real time, right under the original. Even nonstop speech is chunked into 3–5 sentences — all on-device, of course.

Language Off 한국어 English 日本語 中文 English translation

Live demo (sample) — in the real app, the built-in local AI (Qwen3) translates on-device.

Bob live session screen — English and Japanese utterances with real-time Korean translations underneath
Actual app screen — translations appear right under the original as utterances are finalized.
04 · Features

Quiet, but doing everything

An industrial ASR pipeline and on-device AI in a single desktop app. No servers to run, no API keys.

01 · NEW

Real-time fact check

A heuristic gate detects claims with numbers, years, and product names; AI verifies them. Verified · incorrect · uncertain verdicts with evidence — and the correct fact if it's wrong.

02 · NEW

Memory that recalls meetings

After each meeting, decisions, facts, and plans are learned into a knowledge graph. Related topics next time bring back past conclusions with a "Last meeting · date" source.

03

2-pass live transcription

Draft text flows the moment you speak; when a sentence ends, an offline refinement pass finalizes punctuation and timestamps.

04

Speaker-aware transcripts

CAM++ speaker embeddings tell voices apart and save them as profiles — the same person gets the same name next meeting.

05

A suggestion engine that never interrupts

RealtimeGate detects question and decision signals, suppresses duplicates, and respects cooldowns. One "talking point" card, only when it matters.

06

AI that knows your notes

A local md/txt folder is your knowledge base. BM25 search builds note-grounded suggestions with no embedding server.

07

AI summaries & sharing

When the meeting ends, the built-in AI writes the summary. Clipboard copy, email share, Markdown export — one click each.

08

Zero setup & auto-update

First launch installs the transcription engine and AI models automatically; new versions update themselves after sha256 verification.

06 · How it works

How one suggestion reaches your screen

Not every utterance becomes a suggestion. A 3-layer pipeline filters the noise and keeps the context.

  1. 1

    Prefetch

    The moment an utterance is finalized, BM25 searches the knowledge base and stages the top 3 related notes.

  2. 2

    RealtimeGate

    Question, decision, and claim triggers blend with KB relevance scores. Similar suggestions are suppressed; cooldowns respected.

  3. 3

    LLM Synthesis

    Only past the gate does the built-in local LLM (llama.cpp + Qwen3) stream-synthesize a suggestion card.

No LLM setup, no API keys — on first launch Bob prepares its built-in local AI automatically.

07 · Privacy

Your meetings never leave the room

Audio is processed by the built-in transcription engine; summaries and suggestions by the built-in local AI. Recordings, transcripts, speaker profiles, notes — all files on your Mac.

  • No cloud uploads — transcription and AI both on-device
  • Speaker profiles and meeting memory are local files — open or delete them anytime
  • Sessions are JSON + Markdown — open formats, no lock-in
  • Only real-time fact check is opt-in — off by default; when on, minimal checks with your own API key

From your next meeting,
stop taking notes.

Open the DMG, drag, done. No servers, no API keys, no accounts.

Download Bob macOS (Apple Silicon) — auto-updates after install

No setup after install — first launch prepares the speech engine and local AI models automatically, with progress shown in-app. After that it starts instantly.