Meet Live Assist

What a call looks like with it

Four moments, in the order they happen. The "say this" line is one of six markers, and it sits next to the decisions and risks the panel surfaces alongside it.

It does not wait for you to ask

You could type all of this to an agent in a terminal yourself. You would also stop listening while you did it, and in a live call that costs more than the answer is worth. So it reads along, decides on its own when it has something worth your attention, and puts it in the panel before you had thought to ask.

It checks what the room is asserting. "Filters took about two weeks" is the kind of sentence that gets repeated until it becomes a plan. If your tracker says 31 days, you get 31 days and the ticket it came from, before the sizing closes.

One toggle turns suggesting into doing. Turn on 🤖 Auto-create and the moment someone says "raise a ticket for that", the ticket exists, with the link already dropped in the meeting chat if you also turned that on. Same for a planning note or a summary. You authorize once, in the panel, not once per item, and both toggles are off until you flip them.

Not tied to Jira. The tracker is whatever you name in Options, and with no name at all the button says "Draft note" and the assistant produces a plain one. Your agent creates things through the tools your own session already has, so a team on Linear, Asana, Notion or nothing in particular is not a special case here.

What this asks of you

You need an AI coding agent that speaks MCP. Claude Code is the one this is tested with. The extension ships the eyes, ears and hands - capture, panel, local bridge - but not the brain; that is an agent session on your own machine, reading the call through MCP tools. Without one you get a working transcript recorder and an empty advice pane, which is not what the rest of this page shows.

The meeting is other people's conversation too. Turning on captions is invisible to everyone else, unlike recording, which Meet badges. So by default this posts one line into the meeting chat when capture starts, saying an assistant is transcribing locally. You can edit that line or switch it off. Some places require everyone's agreement before a conversation may be recorded or transcribed. That is a call only you can make, and the privacy policy states it up front.

Where we draw the line on interviews and exams, and what happens to a recording you did not ask for: the Q&A.

Whichever agent you already use

This ships the eyes, ears and hands. The brain is whatever agent you already run, and the adapter is plain JSON-RPC over stdio with nothing vendor-specific in it. What matters is not the vendor: it is whether your client can hold a persistent background loop, because MCP is client-pull and nothing here can start your assistant's turn.

• Claude Coderun through a live meeting
• Codextool calls verified here
• Cursorspeaks MCP
• Clinespeaks MCP
• Zedspeaks MCP
• VS Codespeaks MCP

Two of those are claims we can back. Claude Code is the one this has been run through a live meeting with. Codex was checked here: it registers the server, discovers the tools and completes a call against the bridge. The rest speak the protocol, which is the easy half. The hard half is the background loop, and until someone runs one we list them as untested rather than supported. MCP-CLIENTS.md tells you what to check in yours.

Which meeting is yours

Same product, same install. The profile you pick decides which meetings the assistant expects and what it leads with, so it stops guessing at someone else's kind of call.

None of these? generic covers it, or write your own profile as a file - they are about a dozen lines each. A customer call is someone else's conversation as much as yours; the disclosure line above is on by default for a reason.

Where your words actually go

Install it

Three steps, and only the first two need a terminal. Requires Node 20+ and Chrome 116+.

1. Clone and run the installer

   It installs the skill, registers the tools your assistant will call, and prepares the folder your meetings are written to. It changes nothing outside your home directory.

   git clone https://github.com/krystiangw/meet-live-assist-extension.git
   cd meet-live-assist-extension
   ./install.sh

2. Start the bridge server and leave it running

   A zero-dependency Node server, bound to 127.0.0.1. --pair opens a two-minute window for the extension to collect its access token by itself - there is nothing to copy and paste.

   node server/transcript-server.js --pair
   curl -s http://127.0.0.1:8848/health   # → {"ok":true,...}

3. Load the extension, then ask your assistant

   chrome://extensions → enable Developer mode → Load unpacked → pick the folder you just cloned. Pin it and click the icon; the panel pairs itself with the server. Then open Claude Code and ask it to assist your meeting - it loads the meet-live-assist skill, attaches to the live call, and starts answering in the panel.

   Pairing window expired? Run the same --pair command again; against an already-running server it re-opens the window. Or paste the token by hand from <data dir>/.mla-token into the extension's Options - that path still works.

Autostart the server, speak into the call, or transcribe without captions: the optional steps.

Tell us if it is any good

This is early and free, and the only thing that decides whether it grows is whether anyone finds it useful. A sentence about what you expected and what happened is worth more than a filled-in form.

One question worth answering even if nothing else: which profile did you install, and was it the right one? Nothing in the extension reports back, so that answer only exists if you type it.

Who made this

Built by Krystian Gwizdała, for his own calls first. There is no company behind it and no roadmap committee - which is why the answer to "will you add X" is usually just yes or no. More of what I work on →