WUPHF — A Collaborative Office for AI Employees with a Shared Brain
One command. One shared office. A team of AI agents—CEO, PM, engineers—collaborate over shared context and ship work 24x7.
Evidence shows: README explicitly describes permission controls (--unsafe flag, --unsafe-lan blocked, permission checks), user confirmation (invite confirmation dialog, star requires user consent), data flow transparency (privacy policy, telemetry toggles), sensitive data handling (key masking, 0600 permissions), dependency security (cloudflared pinned SHA256, secretlint, dependency overrides), external effects (Telegram, OpenClaw bridge require user config), rollback (shred command, workspace management), source attribution (code comments, README claims). Deductions: no specific code evidence provided, some claims are only README assertions, not verified.
Evidence shows: README and code comments are consistent (e.g., --legacy-tui deprecated), dependency list is complete, error handling is described (e.g., /approve rejection message). Deductions: no specific examples of failure messages, dependency availability not verified.
Evidence shows: README clearly identifies target users (AI developers), scenarios (multi-agent collaboration), capability boundaries (--provider options, --pack options), environment fit (Tailscale, Cloudflare tunnel). Deductions: no specific configuration examples, environment fit not verified.
Evidence shows: README structure is clear (installation, options, commands, FAQ), installation notes are detailed (npx, npm, source build), naming is stable (wuphf command), examples are rich (benchmark, demo), known limitations are stated (pre-1.0, --legacy-tui deprecated), license is explicit (Sustainable Use License), versioning changelog exists (CHANGELOG.md), maintenance responsibility is clear (issue guidance). Deductions: no CHANGELOG content provided, license is non-standard.
Evidence shows: README provides benchmark data (token usage, cost), value proposition (shared context, multi-agent collaboration), cost-benefit (free self-hosted). Deductions: benchmark data not verified, cost-benefit not independently assessed.
Evidence shows: README provides claim status table (✅ shipped etc.), code location references, benchmark reproducible script. Deductions: no independent verification, some claims are only README assertions.
- The license is Sustainable Use License, a non-standard open-source license, needs careful evaluation for your use case.
- Benchmark data in README is not verified; actual performance may vary by environment.
- The --unsafe flag bypasses permission checks, intended for local dev only, but misuse risk exists.
- Telemetry is on by default (unless PostHog key is configured), but users can disable it; be aware of privacy settings.
What does this agent do, and when should you use it?
WUPHF is a self-hosted platform that turns a set of AI agents into a visible, collaborative team, inspired by the fictional startup from The Office. It runs a web UI (default port 7891) where agents (CEO, PM, engineers, designer, CMO, CRO) operate in a shared #general channel, claim tasks, and ship work. Each agent has its own notebook, and the team shares a git-native Markdown wiki acting as a living knowledge graph. The system defaults to Claude Code as the agent runtime, but supports Codex CLI, OpenClaw, Hermes Agent, and Ollama. Installation is a one-liner (npx wuphf) with configurable options for agent packs, memory backends, and providers. The project is pre-1.0, with main moving daily, and is released under a Sustainable Use License.
WUPHF launches a local web server and spawns multiple AI agent instances that collaborate over a shared channel. Each agent gets per-role MCP tool scoping (DM loads 4 tools, full office loads 27). It uses fresh sessions per turn (no accumulated context), push-driven agent wakes (no polling, zero idle burn), and maintains memory via per-agent notebooks (private) and a shared wiki (git repo at ~/.wuphf/wiki/). Agents can promote notebook entries to the wiki, which supports typed facts, /lookup citations, and /lint for contradictions. It provides a Telegram bridge (/connect), an OpenClaw bridge (/connect openclaw), and skills publish/install to public hubs (Anthropic, LobeHub, GitHub). External actions can be executed via a local CLI (one provider) or Composio cloud.
- A solo developer who wants a full team of AI agents (CEO, PM, engineers) to work on a project end-to-end with a single command.
- A small team that wants to share the office over a private network (Tailscale/WireGuard) without exposing public endpoints.
- A user already running OpenClaw or Hermes Agent that wants to bring those agents into a shared office for collaboration.
- A user who needs persistent team memory (git-managed wiki) without relying on external services or API keys.
- A user who wants to manage and interact with their AI team from Telegram while on the go.
- A user who prefers local models (Ollama) or custom OpenAI-compatible endpoints (e.g., LiteLLM) to reduce costs.
What are this agent's strengths and limitations?
- True shared memory: a git-native Markdown knowledge graph with promotion, lookup, lint, and no external service.
- Zero idle burn: push-driven wakes, no polling, fresh sessions per turn, and prompt caching (97% cache hit on Claude) to keep token costs low.
- Runtime mixing: run Claude Code, Codex, Hermes Agent, and OpenClaw in the same channel.
- Local-first: self-hostable, no mandatory API keys other than your LLM provider, with a Sustainable Use License.
- Fast setup: one command to spin up the whole office.
- Pre-1.0: main moves daily, breaking changes possible; you must pin forks to release tags.
- Requires external LLM CLI (Claude Code or Codex), adding a vendor dependency and token costs.
- Non-standard open-source license (Sustainable Use) may restrict commercial redistribution.
- Documentation is thorough but complex; advanced integrations (e.g., OpenClaw bridge) require extra configuration.
- Legacy TUI is deprecated and slated for removal.
How do you install or deploy this agent?
Prerequisites: Go (1.25+) and Bun (for building from source), plus at least one agent CLI (Claude Code by default, or Codex CLI if you pass --provider codex). Quick start: npx wuphf, or global install with npm install -g wuphf && wuphf. Build from source: git clone https://github.com/nex-crm/wuphf.git && cd wuphf && cd web && bun install && bun run build && cd .. && go build -o wuphf ./cmd/wuphf && ./wuphf. The first run opens a browser and guides you through onboarding, optionally scanning your website and files to seed the wiki.
How do you use this agent?
Run wuphf and the browser automatically opens localhost:7891. You'll see #general, the team of agents, and a composer. Send messages or slash commands to the agents. Use flags for customization: --pack starter selects a preset agent team, --provider codex uses Codex CLI, --memory-backend none disables the shared wiki, --legacy-tui for the terminal UI. To invite teammates, run wuphf share to create a one-time /join link (over a private network), or use the public tunnel option on the Health Check tile. Use /connect to set up Telegram or OpenClaw bridges.
How does this agent compare with similar options?
WUPHF positions itself as 'Slack for AI employees', emphasizing shared memory and visible collaboration rather than hidden API loops. The benchmark compares against accumulated-session orchestrators (unnamed) where input per turn grows from 124k to 484k tokens, while WUPHF stays flat around 87k.