WUPHF is an open-source multi-agent framework where CEO, PM, engineers, designer, and CMO agents share a Karpathy-style Markdown/Git wiki brain and ship work autonomously.
A developer released WUPHF on GitHub — an open-source multi-agent orchestration system that spins up role-specific AI agents (CEO, PM, engineers, designer, CMO, CRO) inside a shared browser-based office UI. Agents coordinate via a Git-backed Markdown wiki acting as collective memory, with Claude Code or OpenAI Codex CLI as the underlying agent runtime. The project is pre-1.0, ships via npx in one command, and is built in Go with a tmux-backed dispatch layer. It includes a Nex CRM integration by default, with a --no-nex flag for standalone use.
WUPHF solves the hardest problem in multi-agent systems: shared state. Using a Markdown/Git wiki as a persistent, version-controlled brain means agents have traceable, inspectable memory — not a black-box vector store. The one-command npx install and Go binary mean minimal setup friction. The Claude Code and Codex CLI backends are already in most developers' toolchains.
Clone the repo and swap in your own agent role definition this week — FORKING.md documents how to add custom agent packs, so you can test a specialized 'QA engineer' or 'data analyst' agent against your own codebase in under an hour.
Suggested exercise · not independently tested. Access requirements and costs may vary.
Run: npx wuphf in a new terminal to launch the office in your browser
Once the UI loads, open the generated wiki directory and inspect the Markdown files agents are writing — look at how tasks, decisions, and context are structured across agent roles
Run: cat ARCHITECTURE.md in the cloned repo to see the dispatch layer and how tmux sessions map to individual agent processes
Tags
Sources