Team huddles, transcribed by speaker,
that your agents can read and quote
Codecast huddles transcribe while they run, with the speaker's name on every line. When the huddle ends it gets a title, a summary and action items. Your agents can read it, cite the exact words, and turn what was agreed into tasks.
- MayaRetries are failing on the Stripe webhook again. About one in fifty.
- TheoThat's the backoff. We stop after three tries and Stripe sends bursts.
- PriyaCan we move to exponential with jitter before Friday?
- TheoYes, I'll take it. I'll add a dead letter queue while I'm in there.
- MayaCan the session count how many events we dropped this week?
- PriyaThen I'll write the customer note once we have that list.
- 1. Theo: exponential backoff with jitter and a dead letter queue for webhook retries, before Friday
- 2. Priya: customer note on dropped webhook events, once the list is ready
The first minute of a six minute huddle. One lane per voice, the session that heard it live, and the digest it left behind.
01Each speaker transcribed from their own audio track
Most meeting tools record one mixed stream and guess the speakers afterwards. In Codecast the scribe, one client in the room, holds each person's audio track separately and streams it to speech recognition on its own connection.
A track belongs to one person, so every line carries its speaker without any guessing. If the scribe's laptop closes, another client in the room adopts the same run, and the old one lets go so no word lands twice.
02Where the digest goes when a huddle ends
Every finished huddle with any words leaves one digest: the title, the length, the speakers, the summary and the action items. Where it goes depends on the room.
A channel or DM gets a chat message with the transcript a click away. A huddle in a session's own room wakes that session's agent instead, with the summary and a pointer to the rest.
Stripe webhook retries give up after three tries, and bursts push about one in fifty events past that. The team agreed to move to exponential backoff with jitter before Friday and to tell affected customers once the dropped events are listed.
- 1. Theo: exponential backoff with jitter and a dead letter queue for webhook retries, before Friday
- 2. Priya: customer note on dropped webhook events, once the list is ready
03Reading a call, and agents citing it
The Calls page lists every huddle and voice note you can open, live ones first, with a with video filter. A call's page holds the summary, the action items and the whole transcript under its speakers. Click a line, click another to extend the selection, and Send to agent hands the excerpt to a session with a note on what to do.
Agents read calls the way you read the page. When one quotes a call, the quote shows the speaker and links to those lines on the call's page, and a mention of the call renders as a live pill with its title, length and speakers.
- What did we agree on yesterday's call about the webhook retries?
- Turn the action items from this morning's huddle into tasks, each quoting who took it on.
- Summarize every call this week where billing came up.
Switching retries to exponential backoff with jitter. This was agreed in , and the deadline is Friday:
04Turning action items into tasks
The generated action items are a draft. Ask an agent to turn them into tasks and it reads the summary, checks every item against the transcript (who committed, to what, by when, and whether anyone pushed back later), drops what was only floated and adds what the summary missed.
Each surviving item becomes a task assigned to the person who took it on, with their line quoted in the description and the call linked. Tasks people agreed to reach the human board on their own.
For scripts and agents: the /cast-from-call skill runs this pass, and files each item as a task decided in a meeting.
05Adding an agent session to a huddle
Start a huddle from a session's header, or click + in any huddle and pick A new agent session (Hears the room and answers here). One person is enough to start, because the agent is the second party. You can also point any huddle at another session, a doc or a linked Slack channel, live or once at the end.
Words arrive each time the room goes quiet. A line that names the agent goes through at once, even mid-turn; the rest waits until its turn ends. When it needs to concentrate, the agent can ask for quiet for up to 30 minutes, and whatever was said meanwhile arrives together when the hold ends.
A fed session takes part in the huddle's text chat. The reply it ends a turn with appears there as its own line, cut at 1800 characters with a link to the session, and a line a person types there is relayed in.
On codecast.sh the agent also gets a face: a video participant that says its replies out loud and is marked as an agent on every surface, so nobody takes it for a person. A deployment without a face provider configured keeps agents in the chat.
Delivery acts as the person who added the route, stamped by the server, so a scribe cannot write into sessions only someone else can reach. A guest naming the agent never cuts into its turn; their words arrive as context.
06Recording video and screen shares
Press Record this huddle and the huddle is filmed on the LiveKit server, with a red REC mark everywhere the call shows and a notice for everyone, guests included. Anyone in the call can stop it. Screen shares are kept as their own full resolution files, so code on a shared screen stays legible.
On the call's page the video plays beside the transcript and stays in step with it. An agent can take a frame of any line or second to see what was on screen, and when it cites that moment in a message, the frame renders for anyone who can read the call.
// retry.ts export function nextDelay(attempt: number) { const base = 500 * 2 ** attempt; const jitter = Math.random() * base; return Math.min(base + jitter, MAX_DELAY); } export const MAX_ATTEMPTS = 8;
- Click a line to jump the video to that moment. The line being said lights up as it plays.
- Back to the moment returns to the playing line after you scroll away.
- Link copies a link to that second.
- Screen shares are kept at full resolution, and you can switch the view to them.
07Faces, presence and walkie
The people wall draws the whole team at once, each face sized by how present that person is: active, idle, away or offline, from app check-ins and recent keyboard or mouse input. Click a face for Talk, Ring and Message.
Walkie is push to talk into a DM. The listener hears it live if they are at their desk, everyone else gets the voice recording, and the words show in the DM while they are still being said. It stays one way until somebody presses Join live.
Room access, recording and privacy
A client heartbeats every 15 seconds and a seat older than 45 is ignored, so a closed laptop leaves the room on its own. Everyone joins muted.
An occupied room admits teammates the way a meeting room does. Lock it and a teammate has to knock; anyone inside rings them in.
No account needed. They wait at the door until someone admits them, are told up front when the call is transcribed or recorded, and are marked as a guest on every tile and line.
Only a press of Record starts it. Everyone is told, a beep sounds, and it stops when no teammate is left. The video is readable by whoever may open the room's calls.
A frame of a private call goes to a scratch directory cleared after a day. --share makes a public image only when you ask, and deleting the recording deletes it.
Room rules are enforced in every call mutation and when a media token is minted, because the media server trusts that token.
Commands for scripts and agents
Agents read calls through these commands, and you can script them too. <id> is a short id like cl-42, a full id, or a unique prefix. Every read takes --json.
- $ cast calls
- Calls across your teams, live ones first. Add -n 50 for more history, or --video for only the calls with video.
- $ cast call <id>
- Title, length, speakers, summary, action items, linked sessions, and what was filmed.
- $ cast call <id> --transcript
- The whole transcript under its speakers, each line labeled with the reference that cites it.
- $ cast call <id> 15:25
- Only lines 15 to 25 (15-25 and 15 work too).
- $ cast call cl-42@12:34
- The lines being said at 12 minutes 34 seconds, with that line marked.
- $ cast call <id> --json
- The same as data: segments with seq, speaker_id, speaker_name, text and times. Says whether the call is shared, never its link.
- $ cast call hold <3m|90s|1h|off>
- From a session a live huddle feeds: hold the room's words for a stretch of work. Up to 30 minutes per ask.
- $ cast call snap cl-42:15
- A PNG of the recorded call at the moment line 15 was said. Also @12:34, line ranges, --screen, --composite, --crop, --tiles 2x2. Needs ffmpeg.
Questions and limits
How does it know who said what?+
Is every huddle transcribed?+
Who can read a call?+
Does the agent get the whole transcript dumped into its context?+
What if the call is short or the summary fails?+
What do I need to turn it on?+
Related features
Get started with calls
Install codecast, then have a team admin switch on Calls under Settings, Team, Features. The agents on every member's computer learn to read the team's calls.
Read the calls guidecurl -fsSL codecast.sh/install | shbrew install codecast-sh/tap/codecast or npm install -g @codecast-sh/cli