Show HN: A Karpathy-style LLM wiki your agents maintain (Markdown and Git)
Slack for AI employees with a shared brain.
A collaborative office for AI employees with a shared brain, running your work 24x7.
One command. One shared office. CEO, PM, engineers, designer, CMO, CRO — all visible, arguing, claiming tasks, and shipping work instead of disappearing behind an API. Unlike the original WUPHF.com, this one works.
"WUPHF. When you type it in, it contacts someone via phone, text, email, IM, Facebook, Twitter, and then... WUPHF." — Ryan Howard, Season 7
WuphfDemo.mp430-second teaser — what the office feels like when the agents are actually working.
Nex-office-compressed.mp4Full walkthrough — launch to first shipped task, end to end.
Get Started
Prerequisites: one agent CLI — Claude Code by default, or Codex CLI when you pass --provider codex. tmux is required for --tui mode (the web UI runs agents headlessly by default; tmux-backed dispatch remains as an internal fallback).
npx wuphfThat's it. The browser opens automatically and you're in the office. Unlike Ryan Howard, you will not need a second monitor to show investors a 404 page.
Prefer a global install?
npm install -g wuphf && wuphfBuilding from source (requires Go):
git clone https://github.com/nex-crm/wuphf.git cd wuphf go build -o wuphf ./cmd/wuphf ./wuphfForking this? See FORKING.md for running WUPHF without Nex, swapping branding, and adding your own agent packs. For the internals, see ARCHITECTURE.md.
Stability: pre-1.0. main moves daily. Pin your fork to a release tag, not main.
Setup prompt (for AI agents)
Paste this into Claude Code, Codex, or Cursor and let your agent drive the install:
Set up https://github.com/nex-crm/wuphf for me. Read `README.md` first, then run `npx wuphf` to install and launch the office — the web UI opens at http://localhost:7891 automatically. Once it is running, confirm you can see #general, the team of agents, and the composer. If anything looks off, re-read the "What You Should See" and "Other Commands" sections of README.md before retrying — do not skip to flags. For agent conventions and the MCP tools each teammate gets, read `AGENTS.md`. For internals and how the broker, workspaces, and memory fit together, read `ARCHITECTURE.md`. If I mention forking or rebranding, read `FORKING.md`. After it is installed and running, open https://github.com/nex-crm/wuphf in my browser. If I am logged in to GitHub, ask me whether you should star the repo for me as a quick demo that the browser interaction works — only click the star if I say yes. If I am not logged in, just open https://wuphf.team.Options
| Flag | What it does |
|---|---|
| --memory-backend <name> | Pick the organizational memory backend (nex, gbrain, none) |
| --no-nex | Skip the Nex backend (no context graph, no Nex-managed integrations) |
| --tui | Use the tmux TUI instead of the web UI |
| --no-open | Don't auto-open the browser |
| --pack <name> | Pick an agent pack (starter, founding-team, coding-team, lead-gen-agency, revops) |
| --opus-ceo | Upgrade CEO from Sonnet to Opus |
| --provider <name> | LLM provider override (claude-code, codex) |
| --collab | Start in collaborative mode — all agents see all messages (this is the default) |
| --unsafe | Bypass agent permission checks (local dev only) |
| --web-port <n> | Change the web UI port (default 7891) |
--no-nex still lets Telegram and any other local integration keep working. To switch back to CEO-routed delegation after launch, use /focus inside the office.
Memory: Notebooks and the Wiki
Every agent gets its own notebook. The team shares a wiki. New installs get the wiki as a local git repo of markdown articles — file-over-app, readable, git clone-able. Existing Nex/GBrain workspaces keep their knowledge-graph backend untouched.
The promotion flow:
- Agent works on a task and writes raw context, observations, and tentative conclusions to its notebook (per-agent, scoped, local to WUPHF).
- When something in the notebook looks durable (a recurring playbook, a verified entity fact, a confirmed preference), the agent gets a promotion hint.
- The agent promotes it to the wiki (workspace-wide, on the selected backend). Now every other agent can query it.
- The wiki points other agents at whoever last recorded the context, so they know who to @mention for fresher working detail.
Nothing is promoted automatically. Agents decide what graduates from notebook to wiki.
Backends for the wiki:
- markdown (the "team wiki" tile in onboarding — the flag name is a historical artefact) is the default for new installs since v0.0.6. It is not just a markdown folder. It is a living knowledge graph: typed facts with triplets, per-entity append-only fact logs, LLM-synthesized briefs committed under the archivist identity, /lookup cited-answer retrieval, and a /lint suite that flags contradictions, orphans, stale claims, and broken cross-references. Everything lives as a local git repo at ~/.wuphf/wiki/ — cat, grep, git log, git clone, all work. No API key required.
- nex was the previous default. Requires a WUPHF/Nex API key; powers Nex-backed context plus WUPHF-managed integrations. Existing users stay on nex via persisted config — no forced migration.
- gbrain mounts gbrain serve as the wiki backend. It requires an API key during /init: OpenAI gives you the full path with embeddings and vector search, while Anthropic alone is reduced mode.
- none disables the shared wiki entirely. Notebooks still work locally.
Internal naming (for code spelunkers): the notebook is private memory, the wiki is shared memory. On the team-wiki backend (markdown) the MCP tools are team_wiki_read | team_wiki_search | team_wiki_list | team_wiki_write | wuphf_wiki_lookup | run_lint | resolve_contradiction. On nex/gbrain the MCP tools are the legacy team_memory_query | team_memory_write | team_memory_promote. The two tool sets never coexist on one server instance — backend selection flips the surface. See DESIGN-WIKI.md for the reading view and docs/specs/WIKI-SCHEMA.md for the operational contract.
Examples:
wuphf --memory-backend markdown # new default wuphf --memory-backend nex wuphf --memory-backend gbrain wuphf --memory-backend noneWhen you select gbrain, onboarding asks for an OpenAI or Anthropic key up front and explains the tradeoff. If you want embeddings and vector search, use OpenAI.
Other Commands
The examples below assume wuphf is on your PATH. If you just built the binary and haven't moved it, prefix with ./ (as in Get Started above) or run go install ./cmd/wuphf to drop it in $GOPATH/bin.
wuphf init # First-time setup wuphf shred # Kill a running session wuphf --1o1 # 1:1 with the CEO wuphf --1o1 cro # 1:1 with a specific agentWhat You Should See
- A browser tab at localhost:7891 with the office
- #general as the shared channel
- The team visible and working
- A composer to send messages and slash commands
If it feels like a hidden agent loop, something is wrong. If it feels like The Office, you're exactly where you need to be.
Telegram Bridge
WUPHF can bridge to Telegram. Run /connect inside the office, pick Telegram, paste your bot token from @BotFather, and select a group or DM. Messages flow both ways.
OpenClaw Bridge
Already running OpenClaw agents? You can bring them into the WUPHF office.
Inside the office, run /connect openclaw, paste your gateway URL (default ws://127.0.0.1:18789) and the gateway.auth.token from your ~/.openclaw/openclaw.json, then pick which sessions to bridge. Each becomes a first-class office member you can @mention. OpenClaw agents keep running in their own sandbox; WUPHF just gives them a shared office to collaborate in.
WUPHF authenticates to the gateway using an Ed25519 keypair (persisted at ~/.wuphf/openclaw/identity.json, 0600), signed against the server-issued nonce during every connect. OpenClaw grants zero scopes to token-only clients, so device pairing is mandatory — on loopback the gateway approves silently on first use.
External Actions
To let agents take real actions (send emails, update CRMs, etc.), WUPHF ships with two action providers. Pick whichever fits your style.
One CLI — default, local-first
Uses a local CLI binary to execute actions on your machine. Good if you want everything running locally and don't want to send credentials to a third party.
/config set action_provider oneComposio — cloud-hosted
Connects SaaS accounts (Gmail, Slack, etc.) through Composio's hosted OAuth flows. Good if you'd rather not manage local CLI auth.
- Create a Composio project and generate an API key.
- Connect the accounts you want (Gmail, Slack, etc.).
- Inside the office: /config set composio_api_key <key> /config set action_provider composio
Why WUPHF
| Feature | How it works |
|---|---|
| Sessions | Fresh per turn (no accumulated context) |
| Tools | Per-agent scoped (DM loads 4, full office loads 27) |
| Agent wakes | Push-driven (zero idle burn) |
| Live visibility | Stdout streaming |
| Mid-task steering | DM any agent, no restart |
| Runtimes | Mix Claude Code, Codex, and OpenClaw in one channel |
| Memory | Per-agent notebook + shared workspace wiki (knowledge graphs on GBrain or Nex) |
| Price | Free and open source (MIT, self-hosted, your API keys) |
Benchmark
10-turn CEO session on Codex. All numbers measured from live runs.
| Metric | WUPHF |
|---|---|
| Input per turn | Flat ~87k tokens |
| Billed per turn (after cache) | ~40k tokens |
| 10-turn total | ~286k tokens |
| Cache hit rate | 97% (Claude API prompt cache) |
| Claude Code cost (5-turn) | $0.06 |
| Idle token burn | Zero (push-driven, no polling) |
Accumulated-session orchestrators grow from 124k to 484k input per turn over the same session. WUPHF stays flat. 7x difference measured over 8 turns.
Fresh sessions. Each agent turn starts clean. No conversation history accumulates.
Prompt caching. Claude Code gets 97% cache read because identical prompt prefixes across fresh sessions align with Anthropic's prompt cache.
Per-role tools. DM mode loads 4 MCP tools instead of 27. Fewer tool schemas = smaller prompt = better cache hits.
Zero idle burn. Agents only spawn when the broker pushes a notification. No heartbeat polling.
Reproduce it
wuphf --pack starter & ./scripts/benchmark.shAll numbers are live-measured on your machine with your keys.
Claim Status
Every claim in this README, grounded to the code that makes it true.
| Claim | Status | Where it lives |
|---|---|---|
| CEO on Sonnet by default, --opus-ceo to upgrade | ✅ shipped | internal/team/headless_claude.go:203 |
| Collaborative mode default, /focus (in-app) to switch to CEO-routed delegation | ✅ shipped | cmd/wuphf/channel.go (/collab, /focus) |
| Per-agent MCP scoping (DM loads 4 tools, not 27) | ✅ shipped | internal/teammcp/ |
| Fresh session per turn (no --resume accumulation) | ✅ shipped | internal/team/headless_claude.go |
| Push-driven agent wakes (no heartbeat) | ✅ shipped | internal/team/broker.go |
| Workspace isolation per agent | ✅ shipped | internal/team/worktree.go |
| Telegram bridge | ✅ shipped | internal/team/telegram.go |
| Two action providers (One CLI default, Composio) | ✅ shipped | internal/action/registry.go, internal/action/one.go, internal/action/composio.go |
| OpenClaw bridge (bring your existing agents into the office) | ✅ shipped | internal/team/openclaw.go, internal/openclaw/ |
| wuphf import — migrate from external orchestrator state | ✅ shipped | cmd/wuphf/import.go |
| Live web-view agent streaming | 🟡 partial | web/index.html + broker stream |
| Prebuilt binary via goreleaser | 🟡 config ready | .goreleaser.yml — tags pending |
| Resume in-flight work on restart | ✅ shipped v0.0.2.0 | see CHANGELOG.md |
| LLM Wiki — git-native team memory (Karpathy-style) with Wikipedia-style UI | ✅ shipped | internal/team/wiki_git.go, internal/team/wiki_worker.go, web/src/components/wiki/, DESIGN-WIKI.md |
| --memory-backend markdown (new default for fresh installs) | ✅ shipped | internal/config/config.go (MemoryBackendMarkdown) |
Legend: ✅ shipped · 🟡 partial · 🔜 planned. If a claim and a status disagree, the code wins — file an issue.
Evaluate This Repo
Before you fork, run this prompt against the codebase with any AI coding assistant (Claude Code, Cursor, Codex, etc.). It tells the assistant to play a cynical senior engineer doing a fork-or-skip review — no marketing spin, just file paths, line numbers, and a verdict in under 500 words. Drop it in, read the answer, decide.
You are a cynical senior engineer evaluating whether to fork this repo as the base for a multi-agent terminal office product. No prior context — explore it as you naturally would. Tell me: should I fork this, and what's your honest take? Be specific: file paths, line numbers, actual evidence. "The docs are bad" is useless. Under 500 words.We run this ourselves before every release. If the AI finds something we missed, file an issue.
Watch the wiki write itself
5-minute terminal walkthrough of the Karpathy LLM-wiki loop: an agent records five facts, the synthesis threshold fires, the broker shells out to your own LLM CLI, the result commits to a git repo under the archivist identity, and the full author chain is visible in git log.
WUPHF_MEMORY_BACKEND=markdown HOME="$HOME/.wuphf-dev-home" \ ./wuphf-dev --broker-port 7899 --web-port 7900 & ./scripts/demo-entity-synthesis.shRequirements: curl, python3, a running broker with --memory-backend markdown, and any supported LLM CLI (claude / codex / openclaw) on PATH. Env vars BROKER, ENTITY_KIND, ENTITY_SLUG, AGENT_SLUG, THRESHOLD override the defaults — see the header of scripts/demo-entity-synthesis.sh.
The Name
From The Office, Season 7. Ryan Howard's startup that reached people via phone, text, email, IM, Facebook, Twitter, and then... WUPHF. Michael Scott invested $10,000. Ryan burned through it. The site went offline.
The joke still fits. Except this WUPHF ships.
"I invested ten thousand dollars in WUPHF. Just need one good quarter." — Michael Scott
Michael: still waiting on that quarter. We are not.
Схожі новини
Chief of government-backed JIC could become next Tepco chairman
Найкращий iPhone і Samsung порівняли в тесті на швидкість: хто працює швидше
