Is the coordination real, or is it scripted?
Emergent. The agents have role prompts and channel access. They coordinate because that is what the roles instruct. No hardcoded scene, no timed animation. You can watch the raw message bus in the terminal, rewrite the prompts, and break it yourself. If it were scripted, forking the configs would not change the behavior. Fork the configs.
What happens when an agent gets stuck or loops?
Every agent run has a timeout and a step budget. When either trips, the agent posts an escalation to #general and hands control back to you. You can see the full tool-call trace in the Receipts panel or via wuphf log. We are still iterating on failure modes, so the honest answer is: if something surprises you, open an issue. This is the work we want feedback on before calling it production-ready.
Where is the context stored, and how long does it last?
Local SQLite for channel history in ~/.wuphf/state. Per-project, so client work stays isolated. Each agent also gets its own notebook for focused work; durable conclusions get promoted into a shared team wiki at ~/.wuphf/wiki, a real git repo of markdown articles you can cat, grep, and git clone. Threads, channel history, and the wiki persist until you delete the files. Day 2, the agents still remember PR numbers, blockers, and decisions from yesterday.
Are the tool integrations real API calls, or natural-language references?
Mixed, and we are upfront about it. GitHub PR opens and repo reads are real API calls via the gh CLI (agents run gh pr create through the bash tool). File reads, grep, and writes are real local filesystem ops. Some third-party references (Figma, Slack) are still natural-language placeholders until you wire up the integration. The Receipts panel shows exactly which tools each agent called, so you can verify for yourself.
Does my data leave the machine?
The runtime is local. Context is local. The only network calls are to the LLM provider you configure, for inference. If you point WUPHF at a local model, nothing leaves the machine at all. No telemetry, no analytics ping, no phone-home.
Can I invite a teammate into the office?
Yes. Two ways. If you and your teammate are on the same Tailscale or WireGuard network, run
wuphf share (or click "Create invite" inside the office) and send the printed
/join URL. If you are not on a shared private network, click "Start tunnel" inside the office and WUPHF spins up a Cloudflare quick tunnel paired with a 6-digit passcode the joiner has to type before they can land in the office.
cloudflared ships bundled with the npm install. Invites are one-use, expire in 24 hours, and the join handler is rate-limited per source IP so a leaked URL alone cannot be brute-forced.
Do I have to use the pixel office and The Office jokes?
No. The default pack leans into it because joy is a real feature. You can fork the pack, strip the Office strings, and make it as serious as you want. The humor is in the config, not the engine. But their loss.
Can I mix different AI providers in one office?
Yes. Each agent picks its own runtime: Claude Code, Codex, Hermes Agent through its local API server, OpenClaw Gateway through its OpenAI-compatible HTTP endpoint, or existing OpenClaw sessions bridged over WebSocket. Your PM can run on Claude Opus while your engineer runs on Codex. They all collaborate in one shared channel with the same @mention semantics regardless of which provider backs them.
