An AI Meeting Assistant That Never Hears the Audio

TL;DR Recording a call and summarising it later helps after the meeting, not while someone waits for your answer, and it sends everyone's voice off the PC. Meeting Capture is an AI meeting assistant with no audio input at all: it reads the text Windows Live Captions already shows, so speech recognition stays on the PC, and an answer you can say out loud starts in about 0.8 seconds.

Do this: turn on Windows Live Captions with Microphone + system audio, tell the people on the call you use an AI assistant, and press Ctrl+Alt+Q when the next question starts.

Part 2 ended in a meeting: the question is spoken once, the speaker waits, and the honesty line has to cover audio and the people on the call. Meeting Capture never touches the audio.

The interviewer asks: "How would you test a feature that depends on a third-party payment API?" You press Ctrl+Alt+Q as the question starts. The speaker stops for two seconds, the question sends itself, and the first words of an answer appear in a see-through window.

1. I Built a Lottery App That Refuses to Predict the Jackpot
2. A Spider Crawls Over My Screen and Remembers What It Read
3. An AI Meeting Assistant That Never Hears the Audio (you are here)
4. One Skill Folder That Works in Every Coding Agent
5. An LLM Proxy That Retries a 429 Before Your Client Sees It

No microphone, by design

Meeting Capture is a Windows tray app with no speech recognition of its own. Windows does the listening; the app reads text.

1. Live Captions hears the call. The app starts it and parks it off screen.
2. A reader copies its text through UI Automation, and the app merges the snapshots as Live Captions rewrites its newest words.
3. Your second key press, or a two-second pause, ends the question.
4. The AI answers. It gets the question and about 180 seconds of conversation before it. Claude Haiku 5.5 is the default; Gemini and Grok are options.

An answer, not a summary

Only what was said between the two key presses counts as the question. Everything earlier is background.

The prompt asks for the answer itself: a solution, code, a choice with reasons. scripts/answer-check.js sends tricky captures through the real prompt (no question mark, keywords only, a cut-off sentence) and flags any reply that summarises instead.

The meeting it lost without a word

Live Captions can close on its own. The app started it once, at launch, and never looked again. The commit that fixed it says what that cost:

Live Captions can close on its own (sleep, lock, closed by hand, crash). The app only started it once at launch, so from then on the toolbar kept running but every meeting was lost without a word.

Now the app restarts it when its status drops to off, at most once every 20 seconds.

Try it

git clone https://gitlab.com/agentic-fx/meeting-capture.git
cd meeting-capture
npm install
npm start

In Live Captions, open ⚙ → Caption source and choose Microphone + system audio, so it hears the other people on the call. The app itself never opens a microphone; it only reads caption text.

Each tool so far was written by one coding agent. Your teammates use others: Codex, Cursor, Copilot. The habits that kept these tools honest, like a check that can fail, have to reach every one of them. Part 4, One Skill Folder That Works in Every Coding Agent, carries them in one folder.

Read this with your AI

Read the post "An AI Meeting Assistant That Never Hears the Audio".
1. List what is sent to the AI on each answer, and what never is.
2. Ask me 2 questions to check I understood; wait, then correct me.
3. Draft one line I can send before a meeting to say I use an AI assistant.