promptparty 0.1.3

Cadence sync for teams agent-coding in the same room — one dashboard that says when to PROMPT and when to TALK.

promptparty 🎉

Cadence sync for teams agent-coding in the same room. One shared dashboard that tells everyone when it's time to ⌨️ PROMPT and when it's time to 🗣️ TALK — because the worst part of group agentic coding is everyone heads-down at different moments, waiting on each other, and the conversation dying.

The rhythm promptparty gives you:

1. ⌨️ PROMPT — someone's agent finished (or needs a permission). Read the output, fire the next prompt.
2. 🗣️ TALK — all agents are generating. Hands off the keyboard: discuss what's coming, plan the next move, actually talk to each other.
3. Agents finish → the dashboard flips back to PROMPT (with an optional chime).

Everyone keeps their own terminal, own Claude Code session, own model — promptparty never calls a model and needs no API key. It only watches the states.

Quickstart (60 seconds)

One person starts the server and opens a room (put it on a big screen or a spare browser tab):

uvx promptparty          # or: pipx install promptparty && promptparty

Each teammate runs one command inside the project they're coding in:

uvx promptparty hooks --server http://192.168.1.20:8765 --room a1b2c3 --name Katka

(The dashboard's ⚙️ panel shows this exact command with your room pre-filled.) That installs Claude Code hooks into ./.claude/settings.json that report automatically:

you do	dashboard shows
submit a prompt (UserPromptSubmit)	🤖 agent running — free to talk
agent finishes its turn (Stop)	✅ agent done — read & prompt!
agent asks for permission (Notification)	🙋 needs a decision

Restart your Claude Code session (or reload hooks) and you're live.

Team setup — commit it once, everyone's in

Instead of per-person installs, one person can write a shared, committable config into the repo:

uvx promptparty hooks --shared --server http://SERVER:8765 --room a1b2c3
git add .claude/settings.json && git commit -m "promptparty hooks"

Teammates' names are resolved at runtime from $PROMPTPARTY_NAME (falling back to $USER), and the server/room can be overridden with $PROMPTPARTY_SERVER / $PROMPTPARTY_ROOM — so after git pull, everyone is hooked up with zero setup.

Not using Claude Code?

Any tool can report with one request — wire it into whatever hooks your tool has, or an alias:

curl -X POST http://SERVER:8765/api/ROOM/event \
     -H 'content-type: application/json' \
     -d '{"name":"Katka","state":"generating"}'
# states: prompting | generating | ready | blocked | away

There are also manual buttons at the bottom of the dashboard.

What you get

• 🚦 A giant phase banner: TALK (all agents cooking) / PROMPT (with who the room is waiting on).
• 🃏 A card per person: state, model, how long they've been in it. Blocked agents pulse red so permission prompts stop rotting unnoticed.
• 🔔 Optional chime on phase flips — you don't even need the screen in view.
• 📊 Session stats: talk time vs prompt time and total prompts fired. (Great post-hackathon bragging material.)
• 🧹 Zero config, zero database, zero API keys. Rooms live in memory.

How it works

A single FastAPI process keeps per-room state and pushes it to dashboards over WebSockets. Hook reports are plain HTTP POSTs. The phase rule is one line:

if every non-away member's agent is generating → TALK, otherwise → PROMPT.

That's the whole trick. It works because agent turns are long enough (tens of seconds to minutes) that "all agents running" is a genuine conversational window.

Options

promptparty [serve] [--port 8765] [--host 0.0.0.0] [--min-talk 120]
promptparty hooks --server URL --room SLUG --name YOU [--settings PATH]
promptparty hooks --remove [--settings PATH]

Playing from different networks

The server has no auth — anyone who can reach it can join the room. So the best way to play across networks is to not put it on the public internet at all:

Tailscale (recommended)

Tailscale gives your group a private network between your devices, invisible to the internet and to whatever café Wi-Fi you're on.

1. Everyone installs Tailscale and joins the same tailnet — the host invites the others from the admin console (the free plan is plenty for a party).

2. The host starts the server bound to their Tailscale address only:

   promptparty serve --host $(tailscale ip -4)
   

3. Share the room URL using that address — http://<tailscale-ip>:8765/r/<slug> — and pass the same address as --server when installing hooks. Works from any network; unreachable from everywhere else.

Quick tunnel (zero setup, but public)

No installs for your friends, just a link — good for a one-off session:

1. The host runs the server as usual, then in a second terminal:

   cloudflared tunnel --url http://localhost:8765
   

   (or ngrok http 8765 — same idea). It prints a random public URL like https://lamp-warren-poker-sunset.trycloudflare.com.

2. Share the room as https://<tunnel-url>/r/<slug>.

3. Hooks must point at the tunnel too, not localhost:

   promptparty hooks --server https://<tunnel-url> --room SLUG --name YOU
   

Heads-up: that URL is on the public internet, and anyone who obtains it can watch the room and post fake states. It's presence data, not code — but prefer Tailscale for anything beyond a one-off.

Roadmap

• PreCompact / SubagentStop hook nuances (long multi-agent turns)
• Per-room history graph (talk/prompt timeline you can screenshot)
• Minimum TALK window (--min-talk, default 120 s) — an early-finished agent waits out the window
• Full pomodoro mode: enforced fixed windows
• Adapters for aider / codex / opencode hooks out of the box

PRs welcome — this was born at a hackathon and stays intentionally tiny: one Python file of server, one HTML file of dashboard, no build step.

Security notes

• Anyone with the room URL can view and post states. It's presence data, not code — but run it on a network you trust (see Playing from different networks).
• Rooms are in-memory and vanish on restart.

License

MIT © moudrkat