Voice Standup

A local Codex plugin that uses the OpenAI Realtime API to give you a spoken developer standup from everything you touched in the last 24 hours.

GitHub stars License: MIT Realtime API

Voice Standup is a local-first Codex plugin for an OpenAI Realtime powered desktop companion:

• On startup, automatically gather recent work across local git repos and GitHub activity.
• Ask gpt-realtime-2 for a concise standup.
• Speak the standup with macOS say when available.
• Keep desktop control behind explicit local allowlists.

It is built for builders who wake up, open the machine, and want: “What did I do yesterday, what matters today, what is blocked, and what should I do first?”

Quick Install

curl -fsSL https://raw.githubusercontent.com/Atharva-Kanherkar/voice-standup/main/install.sh | bash

Then add your API key:

cd ~/plugins/voice-standup
cp .env.example .env
open .env

Run your first standup:

set -a; source .env; set +a
npm run standup

Control it with your voice:

npm run control

Manual Setup

git clone https://github.com/Atharva-Kanherkar/voice-standup.git
cd voice-standup
npm install
cp .env.example .env

Set your API key in .env or your shell:

OPENAI_API_KEY="sk-..."

Run:

npm run standup

Voice Standup uses the OpenAI Realtime WebSocket API:

wss://api.openai.com/v1/realtime?model=gpt-realtime-2

Standup Context

By default, Voice Standup scans recent activity from the last 24 hours across:

• local git repos under your home directory and common work folders
• changed, untracked, and recently committed files
• authenticated GitHub events through gh, when available
• lightweight project docs such as README.md, AGENTS.md, and TODO.md

You usually do not need to configure anything. Optional overrides are available for unusual setups:

VOICE_STANDUP_LOOKBACK_HOURS=24
VOICE_STANDUP_MAX_REPOS=12
VOICE_STANDUP_SCAN_ROOTS="/path/to/repos,/another/path"

Codex Plugin Install

This repo includes a Codex plugin manifest at .codex-plugin/plugin.json, an MCP server config at .mcp.json, and a skill at skills/voice-standup/SKILL.md.

If your Codex build supports local plugin marketplaces, the installer writes a local marketplace entry to:

~/.agents/plugins/marketplace.json

That makes the plugin available as a local plugin source. Public search/discovery inside Codex appears to be curated/indexed separately, so GitHub publishing alone may not immediately make it show up next to plugins like Supabase. This repo is structured so it can be submitted or indexed when a public plugin submission path is available.

Startup on macOS

Preview the LaunchAgent:

./scripts/install-launch-agent.sh --mode=control --print

Install voice control at login:

npm run startup:control
launchctl unload "$HOME/Library/LaunchAgents/dev.agentclash.voice-standup.plist" 2>/dev/null || true
launchctl load "$HOME/Library/LaunchAgents/dev.agentclash.voice-standup.plist"

This opens a Terminal window at login and runs npm run control. macOS microphone permissions work much more reliably this way than from a headless background process. Grant Microphone access to Terminal when macOS asks.

Or install one startup standup instead of persistent voice control:

npm run startup:standup
launchctl unload "$HOME/Library/LaunchAgents/dev.agentclash.voice-standup.plist" 2>/dev/null || true
launchctl load "$HOME/Library/LaunchAgents/dev.agentclash.voice-standup.plist"

The LaunchAgent loads .env automatically before starting. Voice-control mode records short commands instead of keeping a continuous hot mic running.

Voice Path

The prototype includes a short voice-input path through sox or macOS ffmpeg.

For one spoken prompt:

npm run voice

For repeated voice control:

npm run control

Try saying:

• "standup"
• "open calendar"
• "open notes"
• "git status"
• "quit"

That records a short utterance, sends PCM audio to the Realtime API, maps it to an allowlisted action, and speaks the result with macOS say. If recording fails, grant microphone permission to your terminal app or install sox with brew install sox. Voice control records short commands instead of keeping a continuous hot mic running.

Safety Model

Desktop actions must be allowlisted in scripts/voice-standup.mjs. Anything that sends messages, deletes files, purchases, installs software, edits system settings, or runs arbitrary shell commands should require an explicit confirmation step.

Why It Is Different

• It does not only summarize one repo.
• It auto-detects recent local project work.
• It blends local git activity with GitHub events.
• It speaks your standup out loud.
• It is intentionally local-first: your API key stays on your machine.

Roadmap

• Continuous push-to-talk mode.
• Wake word support.
• Better microphone device selection UI.
• Calendar and issue tracker context.
• Better wake-at-login setup during install.
• Marketplace-ready screenshots and icons.

Sharing

• Do not commit .env or API keys.
• Do not commit node_modules/.
• If you publish this as a Codex marketplace plugin later, update .codex-plugin/plugin.json with real homepage, repository, logo, and screenshot URLs.

Useful tags: codex-plugin, openai-realtime, realtime-api, voice-agent, developer-productivity, standup-bot, mcp-server, local-first-ai.