Raw capture — raw/agent-standup-github.md

url: https://github.com/Atoll92/agent-standup date_fetched: 2026-10-11

Agent Standup

See who is waiting on whom, and who is blocked — across all your Claude Code sessions — at zero token cost. Agent Standup turns the session files Claude Code already writes 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 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.

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

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.
• 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").

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.

On the shared screen (appears when an agent uses a tool, stays while a command is still running, then ~4 s after the last action) - 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.

Click an agent — its prompt, tool counts and a Recent activity list (last 30 calls with status, duration, and expandable command + output).

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 answer when they get an order ("Running the tests, as usual.", "Grepping like there's no tomorrow."), say goodbye when they leave, and now and then have a little thought; a failing tool call makes the tile shake; confetti when the lead finishes a long job. Whoever receives a report sometimes says thanks ("Got it, thanks."), and an agent that has been thinking for a long time says so ("Hmm, let me think."). When the call has been silent for a while, someone fills the gap the way people do in video calls ("Can you hear me?", "You guys still here?", "Can everyone see my screen?"): after 45 seconds in a live call, about 10 seconds in a replay, at most three times per silence and further and further apart. It is all local (no API), absent from the transcript, and can be turned off in ⚙ → Little lines & animations.

Glance without the panel (VS Code)

The status bar shows live counts for your workspace (needs you, working, your turn) and turns red when an agent may be waiting for you; hover for the sessions, click to open the call on the most urgent one (agentStandup.statusBar).

Context gauge, highlights, loop watch

The header shows how full the lead's context is (ctx 62%, orange then red, with a warning before auto-compaction); tiles and list rows show it from 50%. In Replay, the Highlights bar jumps to the first error, the longest wait, compactions and possible loops. A 🔁 badge marks an agent repeating the exact same call or failing repeatedly. The context window (200k or 1M) is inferred, because transcripts do not state it.

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

Pick one in ⚙ → Language: English, Français, 繁體中文, 简体中文, Español, Deutsch, Português (Brasil), 日本語. The interface (menus, chips, tooltips, recap) is translated; VS Code's command titles, setting descriptions and status bar follow VS Code's own display language. English and French are written by hand; the other languages are AI-drafted and not yet reviewed by native speakers, so corrections are very welcome: each language is one small JSON file (public/locales/<code>.json), see public/locales/README.md to fix or add one (npm run i18n checks the files). The words the agents say stay in the language of your session (French translation of their lines is optional); the little banter lines and the demo exist in English and French.

Your own characters

Replace the drawn characters with your own pixel art. Drop PNG (or GIF/WebP) files, 300 KB at most each, into ~/.agent-standup/characters (⚙ → Characters shows the folder, and in VS Code Open folder opens it; the agentStandup.charactersPath setting changes it): - lead.png (or claude.png) is the lead, you.png is you, any other image becomes a sub-agent (one per agent, repeated if you have fewer images than agents); - name-talk.png next to name.png is an optional open-mouth frame that shows while the character speaks; - 32×32 or 64×64 px, transparent background, facing front, feet at the bottom looks best. The art is shown crisp (no smoothing) on top of the usual scene and desk, and every status, hand and badge still works. Press Reload in the menu to see changes. A starter pack is in examples/characters.

Agents list

On wide windows (900 px+), a fixed list sits left of the call: every agent with its role and one status line, most urgent first (needs you, stalled, working, waiting, done) and the lead pinned on top. It stays readable with dozens of agents, which a grid of tiles does not. Click a name to put that agent in the spotlight (kept for 25 s even when someone else speaks), double-click for details. Toggle it in ⚙ → Agents list.

All your sessions (Sessions view)

The Sessions tab lists every session active in the last 12 hours, across all projects, with one state each: working, needs you?, waiting on sub-agents, your turn, or idle (and how many sub-agents are active). It refreshes every few seconds and reads only the end of each transcript, so thirty parallel sessions stay cheap. Click a card to follow that session in the call.

At a glance

Every tile carries one status: working (with what it is doing), waiting on another agent (who), needs you? (a file edit or write, or a question, has been pending for 15 s+), stalled (waiting on an agent that has been silent for 2 min+), or done. A summary sits in the header, and the tab title shows how many agents need you. The "needs you?" signal is a heuristic: Claude Code does not record permission prompts, so commands (Bash) are not flagged because they can legitimately run for minutes (they get a ⏳ badge after 1 min instead). Measured on real sessions, edits that stay pending for 15 s+ are rare, so the signal is mostly right.

Voices

Four styles in ⚙ (default: Cartoon voices, so everything stays understandable): - Animal Crossing — every agent "talks" in gibberish, one blip per letter, with a sentence melody (up for questions, down for statements) and a register of its own. - Simlish — the same idea with flowing syllables, vowels and little consonant noises. - Cartoon voices — the real words read by your system voice, with a character per agent (chipmunk, deep, robot, wobbly, radio, monster, squeaky). - Natural voices — plain system voices. - Very natural voices — Kokoro-82M, a small open speech model that runs on your machine. Optional: pick it in ⚙ and click Install (about 730 MB, one folder in your user cache, installed with npm; the model is fetched on first use). Until it is ready the system voices keep talking. Mostly English voices, one French voice.

Long lines are subtitled one sentence at a time, in step with the voice.

Handy extras

• Copy as Markdown in the recap: a ready-to-paste standup report (who did what, files, tokens, final report of each agent).
• Notify me when agents finish (⚙): a soft chime, the tab title and, in VS Code, a notification once the team has worked for a while and gone quiet.
• Search the transcript, ambience volume, and a compact layout when the panel is narrow (side bar).

Install

Three ways in — pick one (details in Works with your tools): the VS Code extension (below), the agent-standup command (npx @double_geste/agent-standup --open), or the desktop app (macOS, from the release page). Cursor, Windsurf and VSCodium install the same .vsix.

VS Code

Option 0 — from the Marketplace (easiest)

In VS Code: Extensions view → search Agent Standup → Install. Or: https://marketplace.visualstudio.com/items?itemName=DoubleGeste.agent-standup

Option A — from a GitHub Release

1. Open the repository's Releases page and download agent-standup-<version>.vsix (don't send it by email: Gmail and others block .vsix files).
2. In VS Code: Extensions view → … menu (top right) → Install from VSIX… → pick the file. Or from a terminal: code --install-extension agent-standup-<version>.vsix (macOS: run Shell Command: Install 'code' command in PATH first if code isn't found).
3. Developer: Reload Window, then click the camera icon in the Activity Bar (or run Agent Standup: Open live call) and press Join.

Direct download of the latest version: https://github.com/Atoll92/agent-standup/releases/latest/download/agent-standup.vsix
Or with the GitHub CLI: gh release download --repo Atoll92/agent-standup --pattern "*.vsix", then the code --install-extension line above.

Option B — build it yourself

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

To hack on it instead: open the folder in VS Code and press F5 (launches an Extension Development Host).

Option C — no extension

After cloning the repository (Option B): node server.js and open http://localhost:4747 (see Standalone below). Without cloning: npx @double_geste/agent-standup --open.

The first click on Join is required by browsers to allow sound.

Commands

Settings

Controls

In Replay, a slider and a speed control appear at the bottom, and clicking a transcript line jumps there.

Office music

Turn on Office music in ⚙ and pick a station (FIP, Groove, Jazz, Electro, Reggae, Rock, Nouveautés, Pop) and a volume. The stream is played live from Radio France's public Icecast servers directly in the page; it starts when you press Join (browsers require a click). When any agent speaks, the music drops to ~15 % within a fraction of a second and returns slowly after the last line. Muting (🔊 / M) stops the stream. Music is for your ears only: it is never included in video exports. Stations are for personal listening; nothing is stored or redistributed.

One window talks at a time

If several Agent Standup windows are open on the same machine (sidebar, editor panel, browser tab), only the last one where you pressed Join / Play / 🔊 plays sound; the others stay silent and show 🔈 (click it to take the sound back). This avoids hearing the same call twice, possibly in two languages. A window that stops answering releases the sound after a few seconds.

Voice engines (per OS)

Voices are generated on your machine and played identically in the editor and in the browser:

If no voice is available the agents "mumble" instead. Each agent gets a stable voice from the best ones installed.

Translation (French ↔ English)

Agent messages are in the language of the session. With French selected, each line is translated as it arrives using your local claude CLI (claude -p, Haiku) through two warm background workers: about 3–4 s per line once warm (the very first line takes ~10 s), then spoken. If a translation is not ready in time (7 s live), that line is read in the original language. The menu shows a hint if translation or the French voice is unavailable.

Optional speed-up: set ANTHROPIC_API_KEY in the environment of VS Code / the server and translation calls the Anthropic API directly (~1–2 s).

Privacy & security: everything runs locally. The server listens on 127.0.0.1 only and rejects requests that do not come from localhost (other websites and DNS-rebinding attempts get a 403), so no web page can read your sessions. Translation and jokes send the shortened agent lines to Claude through your own Claude Code login (or your API key if you set one). With English selected and music off, nothing leaves your machine. (Office music connects your browser to icecast.radiofrance.fr while it is on.)

Standalone (no VS Code)

node server.js            # http://localhost:4747   (PORT=… to change)

• http://localhost:4747/ → live, newest session
• http://localhost:4747/?live=0 → replay picker
• ?proj=<project-dir> → limit live to one project folder under ~/.claude/projects

Requires Node 18+.

How it works

Claude Code writes every session to ~/.claude/projects/<project>/<session>.jsonl, and each sub-agent to <session>/subagents/agent-*.jsonl (with a .meta.json linking it to the tool call that created it). Agent Standup merges these into one timeline: orders (the prompt a lead sends to a sub-agent), spoken text, tool calls (with their inputs), the results of those calls (matched by tool-call id, with real durations) and final reports. Each file is parsed once and then only the new lines are read, so even a very large session refreshes in a few milliseconds; responses are compressed. The page polls the server every 0.7 s and plays new events.

The website counts visits anonymously (a 1-pixel file hosted on GitHub per channel: no cookies, no third-party tracker); the app itself has no telemetry.

Troubleshooting

• Nothing happens / "Waiting for a session" — start Claude Code in the open folder; with onlyThisWorkspace the call follows that folder.
• Panel is blank — use Open in browser; if you set a fixed agentStandup.port it may be taken (use 0).
• Two voices at once (e.g. French and English) — update to ≥ 1.8.1: only one window plays sound now. A line that could not be translated in time is read by an English voice, not by a French voice reading English.
• No sound — press Join (or Play) once; check 🔊 isn't muted and Agent voices is on in ⚙.
• French is read in English / hint about translation — Agent Standup looks for Claude Code in your PATH, the usual install folders (Homebrew, npm, nvm, ~/.local/bin, Windows %APPDATA%\npm) and the copy bundled with the official VS Code Claude Code extension. If none is found the app quietly stays in English; set agentStandup.claudePath, or run claude once in a terminal to sign in. If VS Code itself was started from inside a Claude Code session, restart it normally.
• Characters or backgrounds look empty — update to ≥ 1.2.1 (fixed a duplicate SVG id bug).
• Terminal output missing for a command — the call is still running (spinner), or it is an internal call (sub-agent launch / message); outputs are only recorded once a call finishes.
• Pill says OFFLINE — the local server stopped answering (another VS Code window that was serving the port was closed: the extension takes over within a few seconds), or restart node server.js.
• Delay behind the real session — expected ≈ 5–7 s in French (translation), ≈ 2 s in English; the call skips ahead when it falls too far behind.

Limits

• Tool outputs are trimmed (about 1–2 KB each); the shared screen shows the beginning of long outputs.
• Output only exists once a tool call has finished: a running command shows a spinner and its live timer, not partial output.
• Voices on macOS/Windows/Linux differ in quality; Windows needs a French voice installed for French.
• Very large sessions (10 MB+ of transcript) take a second or two to load the first time (about 1.9 s for a 12 MB, 168-sub-agent session on a recent Mac); after that updates take milliseconds.
• Video export and the clip record in real time and need a Chromium-based browser/VS Code. MP4 recording was checked in Chrome; it has not been tested inside VS Code's own panel, which may still produce a .webm.
• "Hide code & file names in videos" does not mask what the agents say (speech bubbles, subtitles, voices).

Credits

Several ideas in 1.23 (settings kept in a file, session stickiness, subtitles that follow the voice, Kokoro voices) started in neondragon's fork of this project. Thank you! Their code was read and rewritten for this codebase. The Traditional Chinese wording was polished with suggestions from a Taiwanese reader on Reddit. Unofficial: not affiliated with Anthropic.

License

MIT

Support the project / work with me

This dev was replaced by his own agents. If Agent Standup made your sessions more fun, you can buy him a coffee ☕ — there is also a ☕ line in the ⚙ menu and at the end of each recap.

Agent Standup is free and MIT licensed. I'm also available for work around AI agents and developer tooling: reach me through my GitHub profile, Atoll92.

Privacy and usage counting

Agent Standup reads local session files and sends nothing anywhere by default. The only optional network uses are: French translation / jokes (your own Claude login, off unless you turn them on), the FIP radio stream (if you turn it on), and the optional usage counter: if you set agentStandup.shareUsage to true (off by default, and ignored when VS Code telemetry is off), the extension makes at most one empty request per day to a 1-pixel file in this repository's releases. There is no identifier and no content; GitHub sees your IP like for any download. It lets the maintainer count active installs (npm run stats).