WUPHF provides a collaborative "office" environment for AI agents (like Claude, Codex, and OpenClaw) to work autonomously, maintaining context and visibility. It targets users building multi-agent AI systems, enabling persistent context and visible progress unlike typical API-bound agents.

How It Works

WUPHF orchestrates multiple AI agents within a shared environment. Core components include per-agent notebooks for private memory and a shared workspace wiki for persistent knowledge. The wiki defaults to a local git repository of markdown files, functioning as a typed knowledge graph, but also supports nex or gbrain (OpenAI/Anthropic) backends. Agents operate via push-driven wakes and fresh sessions per turn, minimizing token usage and idle costs.

Quick Start & Requirements

Installation is straightforward via npx wuphf. Prerequisites include bash and tmux (for TUI mode). LLM providers like Claude Code or Codex CLI are required. Building from source necessitates Go. Links to detailed architecture and forking guides are available within the repository.

Highlighted Details

• LLM Wiki: A unique, git-native knowledge graph backend (defaulting to markdown) for shared agent memory, supporting typed facts, LLM-synthesized briefs, and retrieval without API keys.
• Bridges: Integrates seamlessly with Telegram and existing OpenClaw agents, bringing them into the WUPHF collaborative office.
• Action Providers: Supports external actions through a local "One CLI" or the cloud-hosted "Composio" service.
• Performance: Achieves flat token usage per turn (~40k billed) via prompt caching and fresh sessions, avoiding the exponential growth seen in other orchestrators. Offers zero idle token burn.

Maintenance & Community

Development appears active with daily changes to the main branch, recommending users pin forks to release tags. No specific community or contributor links are detailed in the README.

Licensing & Compatibility

Licensed under MIT, allowing for permissive use, modification, and integration into commercial or closed-source projects.

Limitations & Caveats

The project is pre-1.0, with the main branch being volatile. Live web-view streaming is partially implemented, and prebuilt binaries are pending release tags.