Your work.
Any intelligence.
One operating system.
NWZUP is an independent layer for AI-mediated work. You say what needs doing — it keeps your goals, context, and decisions, plans the execution, routes each step to the right model or agent, verifies the result, and learns from it. The models can change. Your work doesn’t.
Vendor-neutral by design. Claude-first in practice — the compiler, execution engine, and ledger run on the Claude API today.
Interactive sketch: submit a goal and watch NWZUP organize context, plan, providers, execution, and verification as an animated diagram.
Type a goal below — watch NWZUP organize the work around it.
01 The problem
Your work is scattered across a dozen intelligences.
ChatGPT for drafting. Claude for reasoning. A coding agent for the repo. A browser agent for research. Gemini for the long document. Local models for the private stuff.
Every one of them starts from zero. Your goals live in chat threads, your decisions in scattered docs, your progress nowhere at all. Switch tools and the work evaporates — because the work belonged to the tool, not to you.
NWZUP flips the ownership. It is the layer that holds the work itself — goals, context, decisions, artifacts, costs, evaluations — while models, agents, and clouds become interchangeable parts underneath.
From tool-shaped work to work-shaped tools
Today you adapt your work to each model’s window, each agent’s quirks, each cloud’s billing. NWZUP adapts the execution to your work instead: one durable Work State, compiled into plans, routed to whatever intelligence fits each step best — starting with Claude.
02 The Work Execution Graph
One graph holds the whole story.
Every unit of work is a traversable graph — from the person who asked, through the intelligence that executed, to the human who accepted the outcome. Nothing is a black box. Select any node.
Goal
A stated outcome with constraints: what done looks like, what is off-limits, what it may cost.
03 What it does
The machinery under the story.
Work State: your work, kept.
Versioned goals, constraints, decisions, artifacts, accepted and rejected outputs, pending actions, and provenance — persisted across sessions, models, and agents. Close the laptop on Friday; the work is exactly where you left it on Monday, with every decision intact.
In alpha — personal workspaces now; company spaces next.
Work Compiler: plans you can inspect.
High-level requests become structured, inspectable work plans — sequential and parallel steps, dependencies, acceptance criteria, permissions, budgets, and human approval gates. You see what the system intends to do before anything sensitive runs. The compiler runs on the Claude API, so plans read like they were written by a careful chief of staff.
In alpha — running on Claude today.
Provider-neutral execution, Claude-first.
Each step routes to the intelligence that fits it — Claude for reasoning and planning, other models where they win on cost or latency, local models where privacy demands it. Adapters are clean boundaries; no vendor logic leaks into your work. Because the work lives in NWZUP, swapping a model never means starting over.
In alpha — Claude adapter live; OpenAI-compatible and Gemini adapters in progress; local models planned.
Execution Ledger: every run, receipted.
Each real model execution writes an auditable record — provider, model, attempts, latency, tokens, cost, evaluation status, provenance. The ledger already tracks Claude API usage down to the cent, so you always know what the intelligence cost and whether it earned it.
In alpha — Claude usage tracked today.
Evaluate, then recompile.
A generated response is not a successful one. Deterministic checks, structured validation, and your own accept / correct / reject decisions separate output from accomplishment. Over time, the Recompiler compares strategies on real evidence — actual cost, latency, quality, corrective effort — and recommends what to change. No invented savings percentages; only measured runs.
In alpha — evaluation live; Recompiler comparisons planned.
One real workflow, end to end
Submit a research goal. NWZUP keeps the state, compiles the plan on Claude, executes with the ledger recording, evaluates against your criteria, and waits for your verdict. A short loop you can watch — not thirty impressive buttons.
04 Personal & company
Two spaces. One boundary that holds.
Personal work and company work live in separate spaces with separate permissions. Your employer’s admin can govern company data — never your private work.
Personal space
Your goals, your context, your history. Private by default; nothing leaves without your say.
- Work State you own outright
- Bring your own Claude API key or use shared alpha quota
- Export everything, any time
Company space
Shared skills, shared context, team budgets — with administration that respects the boundary.
- Org-owned skills and playbooks
- Admin sees company data, never personal spaces
- Audit trail on every shared execution
05 Architecture
A compiler, not a chatbot.
Under the site is a real system: Work State persists, the Context Compiler assembles what matters, the Work Compiler plans, provider adapters execute, the ledger accounts, evaluation verifies, and the Recompiler learns. Vendor-neutral at the seams; Claude-powered at the core today.
06 Security & privacy
Trust is a feature with a changelog.
What’s enforced today, what’s in alpha, and what’s planned — labeled honestly, because a trust claim you can’t verify is marketing.
Credentials stay server-side
Provider API keys live in server configuration, never in the browser, never in logs. The alpha never asks for your Claude key in chat.
Approval gates hold
Destructive or financially consequential steps pause for explicit human approval. Nothing irreversible runs on autopilot.
Company-grade isolation
Full tenant isolation and SSO arrive with company spaces. Until then, the alpha is personal workspaces only — the boundary is small because the surface is small.
Work that outlives the model.
NWZUP Work OS is in limited alpha with a small group of users. If you do serious work across AI tools and want it to compound instead of evaporate, ask for a seat.