Devotd Github
NewCoding agents

atoll92/agent-standup v1.42.0 — Agent Session Visualization

atoll92/agent-standup is a tool that reads the local session files of multiple AI agents and shows their work status as a team call screen.

Repositoryatoll92/agent-standup ↗

atoll92/agent-standup is a tool that reads the local session files of multiple AI agents and shows their work status as a team call screen.

GitHub description: Your Claude Code, Codex and Mistral Vibe sessions as a live, animated team call: who is working, who waits on whom, who needs you. VS Code extension, terminal command, desktop app. Runs locally, zero tokens.

stars
★ 331
Forks
50
Language
HTML
License
MIT
Latest release
v1.42.0
Read original source ↗View on GitHub ↗

atoll92/agent-standup is a tool that reads the local session files of Claude Code, Codex, Mistral Vibe, and Cowork, and shows work status through a team call animation and board. On a single screen you can check who is working, which agent is waiting on another agent, and who needs a person's help. The latest release is v1.42.0, published under the MIT license. However, the provided information contains no changelog for this release, so what exactly changed in this version cannot be confirmed.

The core is not a model or inference capability but visualization that gathers the progress of multiple agents onto one screen. According to the README, it shows per-agent status, command output, token usage, sub-agents, and screen sharing. Commands running for more than 1 minute and agents that have been silent for more than 3 minutes are flagged through tile status, and the number of failed tool calls is also shown. In screen sharing you can check terminal output and search results, plus file line numbers and diffs. It supports voice, session replay, and video export, and runs as a VS Code extension, a terminal command, or a desktop app. By default it only reads session files and sends no data, but enabling translation or selection features is an exception.

The technical scope is limited to parsing and visualizing local session files, and it does not newly provide any model or inference technology. The README lists many features, but there is little independent verification behind them. Voice synthesis and lipsync depend on local small voice models. So as a control panel for watching the state of multiple sessions it is concrete, but the evidence backing the effect of these features is limited.