Agent Standup
Turns your Claude Code, Codex and Mistral Vibe sessions into a live animated team call showing who works, who waits, and who needs you.
- Source repo
- Atoll92/agent-standup
- Stars
- ★ 382
- Last updated
- 4d ago
- License
- MIT
- Primary language
- HTML
- FA score
- 73/100 · Some gaps
At a glance
- How it runs
- Works with
- Portable with changesCodex · Claude Code
- Cost
- Free, no paid service needed
- Setup effort
- Low · running in minutes
- You'll need
- Typical use
- A developer running several Claude Code or Codex sessions who wants to see at a glance which sub-agent is stuck, waiting, or burning tokens.
- Not a fit if
- Teams that don't use Claude Code, Codex or Mistral Vibe
- Users who need a vetted Windows/Linux desktop build (marked experimental, untested)
- Users who require zero network calls while keeping translation, jokes or radio on
- Website
- agent-standup.com ↗ set on GitHub
- Maintenance
- Commits in the last 30 dayslast commit 2026-11-17latest release v1.47.1 (2026-11-17)+1.9 ★/day over 7 days
- Source review
- 73/100 · Some gaps
What does this agent do, and when should you use it?
Agent Standup is a local visualization layer that reads the session files Claude Code, Codex, Mistral Vibe and Cowork already write and replays them as an animated stand-up call in which each agent is a small character at a desk. Sub-agents join the call when they are handed a task, share their screen to show real terminal output, file diffs, test counts, browser steps, todo lists and artifacts, and leave after reporting. The app parses tool calls and their results, pairs them by tool-call id, and merges everything into one timeline with real durations and token counts, refreshing a local server that only listens on 127.0.0.1. It ships as a VS Code/Open VSX extension, an npx command, a macOS desktop app and two Claude Code plugins, plus a standalone node server.js mode. Replay, a session board, highlights, recap, MP4/WebM export and a 25-second share clip are included; the interface exists in ten languages, with English and French written by hand and the rest AI-drafted and unreviewed.
It tail-reads the JSONL transcripts each tool writes (~/.claude/projects/<project>/<session>.jsonl plus <session>/subagents/agent-*.jsonl, Codex rollout-*.jsonl[.zst], Mistral Vibe's ~/.vibe/logs/session, Cowork's local-agent-mode-sessions), matches each tool call to its result by id, and builds a timeline of prompts, speech, tool calls with real outputs and durations, file reads and diffs, sub-agent hierarchy, token usage and final reports. The page polls the loopback server every 0.4 s (2.5 s when hidden) and animates the new events: speakers take the spotlight, mouths follow actual voice volume, and the shared screen becomes a terminal, editor, browser window, test summary, todo panel or artifact page. It also reads Claude Code's per-session status in ~/.claude/sessions to flag open permission menus and questions as "needs you" in the tab title, status bar and session cards. Recordings can be exported as MP4 (H.264 + AAC) or .webm, or as a ~25-second clip. With translation enabled it sends each line through the local claude CLI (Haiku) into French, Spanish or Brazilian Portuguese before speaking it.
- A developer running several Claude Code or Codex sessions who wants to see at a glance which sub-agent is stuck, waiting, or burning tokens.
- A team recording a demo of how its AI agents work, using the built-in fictional demo or a real session exported with voices and subtitles.
- A VS Code, Cursor or Windsurf user who wants live needs-you/working/your-turn counts in the status bar and a notification when a session starts asking.
- Anyone running long tasks who wants an amber pulse after a command passes one minute, a sleep marker after three quiet minutes, or a repeat-work counter.
- A Linux user driving Claude Code inside tmux who wants to type follow-up lines into the followed session from the call itself.
- A debugging reviewer who replays a past session, jumps to the first error or longest wait via Highlights, and copies a Markdown recap.
Key setup facts
Pulled from the install and usage notes below. The project README has the current values.
- Install / run
npx @double_geste/agent-standup@latest --opennpm i -g @double_geste/agent-standup
- Runtime
- Node.js 18+
- Install notes cover
- macOS · Linux
How do you install or deploy this agent?
VS Code extension (the quickest path): search Agent Standup in the Extensions view, or install from the Marketplace page.
https://marketplace.visualstudio.com/items?itemName=DoubleGeste.agent-standupCursor, Windsurf and VSCodium get the same extension from Open VSX, or install a downloaded VSIX:
code --install-extension agent-standup-<version>.vsixNo editor — run from any terminal (Node 18+):
npx @double_geste/agent-standup@latest --openBuild from source:
git clone https://github.com/Atoll92/agent-standup.git
cd agent-standup
npm run package
code --install-extension agent-standup-*.vsixInstall as a Claude Code plugin:
/plugin install agent-standup --marketplace Atoll92/agent-standup
/plugin install agent-standup-band --marketplace Atoll92/agent-standupmacOS desktop app: download Agent-Standup-<version>-mac-apple-silicon.zip or -mac-intel.zip from the Releases page, unzip and move it to Applications. It is not signed with an Apple developer ID, so on first launch use System Settings → Privacy & Security → Open Anyway (or right-click → Open).
How do you use this agent?
After installing, click the Agent Standup icon in the Activity Bar or run Agent Standup: Open live call. With no session yet, press ▶ Watch the demo. Browsers block autoplay, so the call starts silent and your first click or key turns the sound on.
Standalone server:
node server.jsThen open http://localhost:4747 (live, newest session) or http://localhost:4747/?live=0 (replay picker).
Claude Code status line:
npm i -g @double_geste/agent-standup{ "statusLine": { "type": "command", "command": "agent-standup statusline" } }Key settings: agentStandup.codex, agentStandup.vibe, agentStandup.reply, agentStandup.onlyThisWorkspace, agentStandup.charactersPath, agentStandup.leadName. CLI equivalents: --codex off/DIR, --vibe off/DIR, --cowork off/DIR, --projects DIR, --reply. The reply field needs Claude Code running inside tmux on Linux.
What are this agent's strengths and limitations?
- Reads only local session files and, with defaults, calls no model, so it costs zero tokens; the server binds to 127.0.0.1 and returns 403 to non-localhost requests.
- One window covers Claude Code, Codex, Mistral Vibe and Cowork sessions, including each tool's sub-agent hierarchy.
- Shows verified execution detail rather than summaries: real command output, real durations, recorded diffs, real test pass/fail counts and recorded token usage.
- Many delivery paths — VS Code/Open VSX extension, npx command, macOS desktop app, two Claude Code plugins, and a standalone node server.js.
- Replay, session board, highlights, Markdown recap, MP4/WebM export and a 25-second clip, plus an optional hide-code mode for sharing.
- Optional extras pull in dependencies or network traffic: natural voices need a ~750 MB Kokoro-82M download, and translation sends agent lines to Anthropic on your Claude Code plan.
- Fidelity depends on each upstream format: compressed Codex sessions need Node 22.15+, and Mistral Vibe's two engines behave differently — the legacy harness records no timestamps and never signals waiting.
- Windows and Linux desktop builds are labelled experimental and untested by hand; Cowork on Windows and Linux paths are likewise untested.
- The reply field is Linux + tmux only, and once enabled anyone who can run commands as you can type into your sessions — the README states this risk plainly.
- Non-English/French interface translations are AI-drafted and unreviewed, and the Halloween character art is not part of the open-source code.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Agent Standup This agent | 73 · Some gaps | IDE extensionFree | ★ 382 | 4d ago | HTML | Codex · Claude Code |
| Claude DevTools | 73 · Some gaps | Desktop appFree | ★ 4k | 1mo ago | TypeScript | Claude Code |
| Sidekick Agent Hub | 73 · Some gaps | — | ★ 91 | 2d ago | TypeScript | Codex · Claude Code |
| ccglass | 72 · Some gaps | CLIFree + model costs | ★ 989 | 1mo ago | JavaScript | Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
Evidence shows: desktop main.js uses contextIsolation, sandbox, nodeIntegration:false, a permission allowlist of only notifications and clipboard-sanitized-write, and https-only external link allowlists; the reply feature is off by default, Linux-only, tmux-gated, token-protected and idle-checked, with its risk stated explicitly. Data-flow transparency is strong: README and CLAUDE.md repeatedly state it only reads local session files, when it calls Claude (translation/jokes), and the env switches. Deductions: reply, once enabled, can inject input into sessions — a high-privilege capability guarded mainly by documentation; dependency security shows only puppeteer-core as a devDependency and no runtime deps, but no lockfile or audit evidence; rollback has no version-revert/undo mechanism described beyond git and the release process; source attribution has MIT and trademark notices, but publisher identity is unverified.
Evidence shows: CI runs syntax checks, version consistency, npm test, i18n and a CLI smoke test; test files cover accounts, board and whiteboard scenarios; failure messages are given in SKILL.md (e.g. Unknown command, No session to credit). Deductions: no lockfile evidence, so dependency availability relies on npx pulling a pinned version; tests are statically visible but not executed here; some platforms (Windows/Linux Cowork) are self-described as untested.
Evidence shows: targets VS Code, Cursor, Windsurf, terminal and desktop app, supports Claude Code/Codex/Vibe/Cowork sources, and SKILL.md states trigger conditions and boundaries (when to use, when not, how to fall back if commands cannot run). Deductions: trigger precision depends on description text, and several skills (status/report/credits) have similar descriptions that may confuse; environment fit has platform limits (reply Linux-only, compressed sessions need Node 22.15).
Evidence shows: CLAUDE.md gives a clear repository map and run/test notes; README install paths are complete (Marketplace, Open VSX, npm, macOS); known limitations are detailed (Vibe engine differences, untested Cowork platforms, video export formats); LICENSE is MIT. Deductions: no CHANGELOG content sample, version consistency is CI-checked but not shown; maintenance responsibility rests with an individual author and the publisher is unverified; naming stability is muddied by coexisting names (agent-standup / double-geste / DoubleGeste).
Evidence shows: output usability is strong (live call, board, token counts, video export, clip, timesheet/credits skills producing directly usable text), and marginal value lies in multi-session aggregation and visualization. Deductions: the core value is largely presentational with limited gain to actual coding throughput; on cost-benefit, zero tokens is the selling point, but the speech model needs a one-time install and translation/jokes call Claude, so benefit depends on the user running many parallel sessions.
Evidence shows: README claims cross-check against code/tests (e.g. reply's 404, permission allowlist, CI steps), and SKILL.md explicitly separates data from instructions. Deductions: many functional claims (e.g. tested against Vibe 2.26.1, Cowork tested on macOS) cannot be verified from the static sample; some numbers and behaviors are author assertions without traceable test references; fact/inference separation is decent in docs but not systematic.
- With reply enabled, input can be injected into a tmux-hosted Claude Code session; the docs themselves note anyone who can run commands as you (including a sandboxed agent reaching localhost) could exploit it — keep it off on shared machines.
- Translation and jokes call Claude (the user's own claude or ANTHROPIC_API_KEY), a conditional caveat to the 'zero tokens' claim; disable as needed.
- Publisher identity is unverified; verify the Marketplace/Open VSX publisher against the repository before installing.
- Some platforms (Cowork on Windows/Linux, desktop builds for Windows/Linux) are self-described as untested, so cross-platform behavior is uncertain.
- No lockfile or dependency-audit evidence; npx pulling a pinned version still depends on upstream availability.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
v1.47.1 1.47.1 — The call asks for a rating itself
Release notes excerpt
- In VS Code, the call asks for a rating itself. When a rating is due (from the third day you use it), the call asks once, in a small line by the ⚙ menu, at the end of a run you watched: the lead at work for two minutes or more, its turn done. Before, a notification asked, and VS Code tucks those into the bell after a few seconds. “⭐ Rate it” opens the store page your editor installs from, “No thanks” ends it, and anything else asks again a week later, one last time. If you have not opened the call for a week, the notification still asks. Nothing is sent. …View on GitHub ↗ -
v1.47.0 1.47.0 — The call in German, Danish, Japanese, Korean or Chinese
Release notes excerpt
- The call in German, Danish, Japanese, Korean or Chinese. With one of them as the interface language, the call asks once: “Hold the call in this language?” Until you say yes it stays in English, as before. After yes, what the agents say is translated by your own Claude Code, and the call's own lines (answers to orders, goodbyes, small talk) are written in that language: Chinese in Simplified, and in Traditional as used in Taiwan for the Taiwanese interface. …View on GitHub ↗ -
v1.46.0 1.46.0 — Claude Whiteboards and Sheets on the shared screen
Release notes excerpt
- Claude Whiteboards and Sheets on the shared screen. When a session fills a Claude Whiteboard or a Claude Sheet, the call shows it on the shared screen as each write leaves it: the board with its sections, blocks and arrows, the Sheet as a grid with its tabs and what its formulas came to. It is drawn from the session file alone; nothing is fetched from claude.ai. …View on GitHub ↗
FAQ
Will it spend my model quota?
Which tools does it support, and is setup required?
Why is my session missing from the call?
Do exported videos leak code?
What runtime does it need?
Is Agent Standup free?
Is Agent Standup open source?
Is Agent Standup safe to use?
Is Agent Standup still maintained?
Who makes Agent Standup?
What is Agent Standup's official website?
Thanks — we use this to improve the page.