Atoll92/agent-standup

Agent Standup

See who is waiting on whom, and who is blocked — across all your Claude Code and Codex sessions — at zero token cost. Agent Standup turns the session files Claude Code and Codex already write into a live team call (and a board of every session): it only reads local files and never calls Claude unless you turn on French translation or the jokes.

Your Claude Code or Codex session as a live team call. The lead gives orders, sub-agents join the call, work at their desks, share their screen and report back — with voices, expressions and a view of the real work as it happens.

Runs as a VS Code extension, a terminal command, or a desktop app. Try it in 30 seconds, no install and no editor (Node 18+): npx @double_geste/agent-standup --open, then ⚙ → Source → Demo.

→ Website & download: https://atoll92.github.io/agent-standup/

_A fictional team fixing a slow login page — the built-in demo (⚙ → Source → Demo). Watch the full 1-minute video._

%20dev%20a%20coffee-ffdd00?logo=buymeacoffee&logoColor=black)

Unofficial community project, not affiliated with or endorsed by Anthropic or OpenAI. "Claude" is a trademark of Anthropic; "Codex" is a trademark of OpenAI.

What you get

• Token usage per agent — a ⚡ counter in the header and on each tile shows how many tokens each agent used (input + output + cache writes; cache reads shown separately in the tooltip), counted once per message as recorded by Claude Code. Handy to see which sub-agent is the expensive one. It is read from the session files, not an estimate of your bill.
• Stuck or quiet agents — in the live call, a tile pulses amber with ⏳ when a command has been running for over a minute, shows 💤 when a working agent has gone quiet for 3 minutes, and a red ✗ counter for failed tool calls. The browser tab title shows how many agents are working.
• Built-in demo — no session yet? ⚙ → Source → Demo plays a fictional team (no data of yours).
• Live call — follows your newest Claude Code session in real time. Each agent is a small character at its desk; the one who speaks gets the spotlight and keeps it until someone else talks. Agents that talk back-to-back talk over each other, like a real meeting.
• Real voices, lip-sync — lines are spoken aloud; mouths follow the actual volume of the voice. For the best experience, switch to the very natural voices (see Voices).
• Live screen sharing with the real output — when an agent works, its screen fills the stage (presenter's camera in the corner, "sharing" badge): a terminal with the command history and the actual output of each command, search results and tool calls; or an editor with the file it reads (real content and line numbers) or the exact diff it applies. See What you see for each agent below.
• Agents come and go — a sub-agent joins when it is given a task and leaves the call 2 minutes (playback time) after its report, with a dissolve animation. It comes back if it gets a new task.
• Sub-agent tree — the full hierarchy (you → Claude → sub-agents → their sub-agents) with status, tool counts and the current action.
• Click any agent — its prompt, tool usage and reports.
• Replay — pick any past session, scrub through it, see a timeline of who worked when, and get a recap (agents, tool calls, commands, files changed, busiest agents).
• Video export — ⚙ → Tools → 🎞 Export video records the call at 1280×720 with voices and subtitles, as an MP4 (H.264 + AAC) when the browser can record it, and as a .webm when it cannot. X, phones and most chat apps refuse .webm.
• Clip to share — ⚙ → Tools → 🎬 Clip to share (25 s) records about 25 seconds, saved as agent-standup-clip.mp4 (.webm when the browser cannot record MP4). In the middle of a replay it starts at the current position; otherwise one spoken line before the first error or loop of the session, and failing that at the moment the lead first hands out work. The line being said when time is up is finished rather than cut, so the length varies by a few seconds. The clip carries the site address (atoll92.github.io/agent-standup) in the top right corner and ends with a short end card; the full export has neither.
• Hide code & file names in videos — ⚙ → Tools, off by default, for the full export and the clip. Commands, their output, file contents, diffs and search results on the shared screen are drawn as grey bars of the same length; file names are removed from the window title and from the activity tags on tiles (the tool's name, e.g. Bash or Read, stays); the session title is replaced by "Agent Standup"; pictures (screenshots, images) become an empty frame. What the agents say is not masked: speech bubbles, subtitles and voices are kept as they are, so check the clip before posting if agents quote code or names aloud.
• French or English — agent text is translated on the fly (see below).
• Office music — optional live web radio (FIP and its themed stations) that fades down smoothly while an agent talks and fades back up when the room goes quiet.
• Office ambience, optional (off by default) and fully synthesised: keyboards that type when agents run commands, clicks when they read, a coffee machine, a phone buzzing, a far-away door… (⚙ → Office ambience). Plus small reactions and jokes between agents ("episode mode").

Codex sessions too

Sessions of OpenAI’s Codex (the desktop app and the codex command) appear next to your Claude Code sessions: in the session list (project “Codex”), on the board, and in the live call when one of them is the most recent. Nothing to set up: the app reads the files Codex already writes in ~/.codex/sessions (or $CODEX_HOME), and never writes there.

• What is shown: what you typed, what the agent says, each command with its real output and whether it failed, files it created or changed, web searches, MCP calls, the questions it asks you, its reasoning summaries as inner thoughts, token counts, and the end of each turn. The lead is called Codex.
• Not yet: Codex sub-agents (a Codex session is the lead and you), and pictures. The short reviews Codex runs on its own before a command are not shown as sessions.
• In VS Code the call follows the Codex sessions started in the folder you have open, like it does for Claude Code.
• Turn it off or point it elsewhere: in VS Code, the setting agentStandup.codex; anywhere else AGENT_STANDUP_CODEX=off, or AGENT_STANDUP_CODEX=/path/to/codex-home (command line: --codex off or --codex DIR). With --projects DIR, Codex sessions are only read if you also pass --codex.

What you see for each agent

On its tile

• Activity tag (top right) — what the agent is doing right now: Reading App.jsx, Editing foo.ts, Writing …, Running npm test (or the command's own description), Searching “pattern”, Listing **/*.js, Fetching host, Updating todo list; then ✓ done when it reports.
• Ticker (bottom) — the current tool call and target, with a spinner and a running counter while it executes, then ✓ / ✗ and its real duration.
• Status border: orange while working, spotlight while speaking, the tile fades and dissolves 2 minutes after its report.

• Terminal — ❯ command (with its description), then its real output (first lines, errors in red, “… +N lines”), ● Grep(…) / ● Glob(…) with matches, a spinner for what is still running. Footer: ✓ Bash · 1.2 s or ⠋ Bash · running 3.4 s.
• Editor — for Read: the file with its real content and line numbers (read-only); for Edit: the removed/added lines; for Write: the new content, with the tool's result line.
• Pictures — when a tool gives the agent a picture (a screenshot it takes, an image file it reads: PNG, JPEG, GIF, WebP), the picture is shown on the shared screen, in a taller window, and stays up for several seconds (8 in a live call) whatever the agent does next. It is also in the agent's Recent activity and in exported videos. Pictures are read from the session file itself, never from your disk.
• Task list — if the agent keeps a todo list, a side panel shows it live (☐ pending, ◐ in progress, ☑ done).
• The presenter's camera lip-syncs when they speak; the subtitle bar shows what they say.

Outputs are trimmed (about 1–2 KB each) to keep the call fast. Results of internal calls (sub-agent launches, messages) are not shown.

Screenshots

| Real commands, real output | The exact change | Who asked whom | |---|---|---| | | | |

| Task list of an agent | End-of-episode recap | In French | |---|---|---| | | | |

Personality, without noise

Sub-agents also talk to each other, from what really happened: a new agent asks a question and the lead answers, a report gets a reaction that fits (all good, errors along the way, it took long) and sometimes a word from a third agent, an agent that fails three times in a row gets a “you okay?”, and every few minutes the lead calls a quick round. ⚙ → Display → Chatter between agents: Quiet (only the real exchanges) / Normal (default) / Lively. When the lead finishes, it reads the part of its last message that the live reading had cut short, so you know what was done.

Glance without the panel (VS Code)

Context gauge, highlights, loop watch

Works with your tools

Other agent tools than Claude Code (Codex CLI, Gemini CLI, OpenCode…) are researched in ROADMAP.md: their log formats are mostly undocumented, so no adapter is shipped yet.

Languages

git clone https://github.com/Atoll92/agent-standup.git
cd agent-standup
npm run package                       # needs Node 18+; produces agent-standup-.vsix
code --install-extension agent-standup-*.vsix