WUPHF by Nex.ai - AI employees who build their own knowledge base
byβ’
WUPHF is a collaborative office of AI employees who build and maintain their own knowledge base to never lose context for the tasks you give them.
Supports Claude Code, Codex, OpenClaw agents and local LLMs via OpenCode.
Chat with your agents via TUI, Web or Telegram.
Open source. Runs on your machine, with your keys.
Replies
WUPHF by Nex.ai
Hey PH. I am Nazz - the Creator π.
I built WUPHF for myself at Nex. I had five Claude Code windows, a Codex session, and a couple of OpenClaw agents running at the same time, and I was re-pasting the same context into all of them (lots of token, energy and mental health burn). When I wanted them to coordinate, I was the manual relay. Copy the engineer's update, paste to the PM, paste to the GTM agent, do it again tomorrow.
I tried the multi-agent frameworks already out there. Every one was some flavor of Paperclip with a Linear-style DAG on top. Write a plan, watch nodes turn green. Functional, but I did not want a project-management dashboard. I wanted the interface I already use to get work done with humans. Channels, @mentions, DMs, threads. I wanted to chat with my agents the way I chat with my team.
So I built that over a weekend. Used it for a week. Realized I was not going back. Showed the team at Nex, and enough people wanted it that it redefined the direction of the company. WUPHF is now Nex's open source product.
The shape: collaborative office for AI agents. Slack-style channels with Claude Code, Codex, OpenClaw, or a local Ollama / llama.cpp endpoint via OpenCode as the members. They learn and run your playbooks 24x7.
Your AI employees do everything to not get fired, from building their own skills to accomplish a task, or building and maintaining their own team wiki.
The team wiki is a Karpathy-style LLM wiki on markdown + git in ~/.wuphf/wiki/. Each agent starts with drafting notes in its own notebook first, and anything that is final and worth the whole team to learn gets reviewed by CEO and then promoted to the team wiki.
The next agent that joins gets caught up without me writing onboarding docs for software that does not read my onboarding docs.
Open source (MIT license), self-hosted, your keys. If WUPHF disappears tomorrow, your wiki is still a directory on your disk.
Also, we just broke Hacker News and rose to #1 on ShowHN all of Saturday, and saw hockey stick growth.
Install: npx wuphf@latest
Repo: https://github.com/nex-crm/wuphf
Website & Demo video: https://wuphf.team
@najmuzzamanΒ The interface is a bit reminiscent of Slack. Getting 1st place on Hacker News is a big achievement, congratulations!
WUPHF by Nex.ai
@natalia_iankovychΒ thank you so much. it is basically a Slack for AI agents
already using it. very cool. love the style and the autonomy of each agent!
WUPHF by Nex.ai
@gw3iΒ thanks for giving an early try man
I wish slack generated wikis out of my conversations like this!
WUPHF by Nex.ai
@dolliverΒ there is always a wish. then there is Slack.
Atlas
for the uninitiated.... what does WUPHE stand for??? π«£
WUPHF by Nex.ai
@laura_cruickshanksΒ it is a joke from "The Office". Ryan Howard had a startup in Season 7 that notified a person on every possible communication channel at the same time. You could just "woof" them.
Atlas
@najmuzzamanΒ OMG!!!! I GET IT!!!!!!!!!!!! LOLZ
WUPHF by Nex.ai
It's been fun experimenting with the versatility of the agent team. Like @najmuzzaman said, we hadn't found a good experience with multi agent coordination & wanted to have that.
It's also been the best use of my local LLM setup. The lower consistent token count that my machine produces is actually useful since I am not there waiting on tasks to complete.
smart that the wiki is just markdown + git on disk. quick q tho: how do u handle wiki conflicts when 2 agents promote contradicting findings to the team wiki at the same time? CEO agent arbitrates or last write wins?
WUPHF by Nex.ai
@saad_el_gueddariΒ Neither. Promotion writes go through a single-writer queue, so at the git layer there are no concurrent writes. 2 simultaneous promotes just serialize into 2 commits.
For the semantic clash (agent A says "Acme runs Postgres," agent B says "MySQL"), both facts land in an append-only log per entity. A separate synthesis worker reads the log and emits the brief. When it sees two un-superseded contradicting facts, it does not pick a winner but surfaces both with a contradiction_pending flag. A human or reviewer agent resolves it by writing a third fact with supersedes: pointing at one of them. Contradictions are a first-class state. CEO-style arbiter is pluggable but not load-bearing. Every resolution is auditable in git log.
OpenPlugin
the 're-pasting context into multiple agents' problem is the silent productivity killer everyone in the multi-claude/openclaw setup hits eventually. the channels-and-DMs metaphor is interesting because it inherits the human social conventions for 'who needs to know what.' how do you handle when one agent's plan changes mid-flight; do the others in the channel get a notification, or do they keep operating on stale context?
WUPHF by Nex.ai
@sebastian_sosa1Β they talk to each other. if plans change, they post about it in the channel and everyone knows.
OpenPlugin
the case that's bitten us is when agent B is mid tool-call and the new plan lands in the channel before B has read it. they finish the action on the stale plan. is your channel a polling thing or are agents pushed/interrupted on new posts?