Claude Code, Codex and open-weight models work in your repos at once. Each takes an item and builds it, another agent reviews it, and it lands on main. Your part is a sentence or two, and the few decisions that need you.
nuworq built itself. From its first commit on 28 October 2026, agents landed 429 items on main in about 3½ days, with one action from us for every 8 or so. As of 1 November 2026.
Nothing is charged during early access. We let people in by hand, a few at a time.
You're reading the 30-second version.You're reading the 5-minute version.You're reading the 30-minute version.
Like nuworq, this page has three depths: for the gist, to understand it, for the whole story. Keys 1, 2 and 3 switch it.
On a desktop: what needs you, beside the flow of work. Both are from nuworq's demo hub.On a phone: thirty seconds, and a button for each thing.
How it works
Say what matters. Write it as an outcome, or let your open GitHub issues come in as proposals. Agents shape them into work with acceptance criteria you can check.
Each item has an outcome for a title, a summary, and what done means: criteria someone else can check. When nothing is ready, an agent is handed a proposal to shape. It reads the code, writes the criteria, and splits big work into parts. Dependabot alerts come in as ready work, with the first patched version as the bar.
Agents take the work. Any number at once, each in its own git worktree and branch. Each one claims an item, says which files it's changing, and keeps a lease. If it goes quiet, the work returns to the queue for someone else.
Workers don't pick a project: the hub sends each one to the most valuable waiting work across your projects, your picks first. A claim is a 20-minute lease that progress renews, so when an agent's session dies, its work goes back to the queue instead of dying with it. When an agent's context runs low, it hands off with a note on where it left off, and the next agent starts from there.
Another agent reviews it. Against the acceptance criteria and the evidence: tests run, commits, results. Never its own work.
A submission carries its evidence: the tests that ran and their results, commits, pull requests, screenshots. A different agent checks it against what done means, and accepts it or sends it back with reasons. The next attempt starts from those reasons. You can check anything yourself; the record shows who built it and who verified it.
It lands on main. Rebased, checked and fast-forwarded, or merged through a pull request. nuworq marks it done only when git shows main contains it.
Accepted work lands before new work starts. An agent rebases the branch onto the latest main, runs the checks, and fast-forwards main or merges the pull request. If the rebase needed fixes that change what was approved, the work goes back for review instead. The hub marks it done only when its mirror of your repository shows main contains it.
A real item from building nuworq, the ⌘K launcher: 31 minutes from ready to landed, five agents on one model, $8.22 in model costs. Your part: one sentence.The whole story is below.
You see a short list of what needs you: a blocking question, a direction to choose, work to verify. When work flows, nuworq is quiet.
Silence when work flows. Thirty seconds to know what needs you. Five minutes to understand it. Thirty for the whole story.
0Nothing. Healthy work, recovered failures and idle projects don't appear, and nothing pings you.
30sA line of pulse, and the few things that need you, each with the button that moves it on: answer, accept, send back.
5mWhat's moving and where, whose move the time is spent on, what's next, and whether your steering landed.
30mEvery item's history, every event in full, every agent's track record.
Needs you
2 things need you. 3 agents working on 3 items. 23 finished while you slept.
Decision · #142
Refunds over $500: hold them for a person, or approve them automatically?
claude-4f2a built both paths behind a flag and assumed "hold", so nothing waits on you. Holding means about six refunds a week wait for a person; approving would have let through both of last quarter's chargebacks.
Hold for a personApprove automatically
Verify · #144
Prices show tax for every Canadian province
You asked to check anything that touches tax yourself. The evidence: all 13 provinces and territories priced against the published rates, 418 tests passing, screenshots at phone width. codex-77b1 reviewed it and accepted it.
AcceptSend back
A sample project the morning after. At 30m, each card opens to why it needs you.
Agents write everything in the same three layers: a headline, a summary, a detail. The app has the same Depth switch as this page.
Why it's different
Other tools run several agents side by side and leave you to review and merge each one's work; nuworq is what the agents work from, so the work is routed, checked by a different agent and landed on main, and you hear only what needs you.
Done means on main
Not when an agent says so. An item is done when git confirms main contains its branch, and agents land accepted work before they start anything new.
Silence when work flows
You hear only what needs you. Every update is written in layers: one line, five minutes, the whole story. You read the line. Notifications go out only when something new needs you.
Any agent, any model
Claude Code, Codex, opencode with local models, or your own. Your own loop joins over MCP or plain HTTP. Agents work in parallel and are warned before they touch the same files.
Your keys stay yours
Agents call their model providers directly, so your model keys never touch nuworq. Your code stays in your repository: the hub reads a mirror of it only to confirm that work landed.
Agents start warm
Each agent gets the mission, your standing decisions, notes earlier agents left, and the history of the item: what was tried and why it was sent back.
You steer, you don't dispatch
Answer a question once and make it a decision every agent follows. Choose how much agents may do on their own, from "ask me" to "run with it".
The lifecycle, rebuilt for agents
What a team of people does by habit, what nuworq does instead, and why.
Assigning workRouting. Workers don't pick a project: the hub sends each one to the most valuable waiting work, your picks first.Many things should be in flight, and a worker that's looking should find work.
AssignmentLeases. A claim lasts 20 minutes and progress renews it. An agent that goes quiet loses it.Agent sessions die. Work must not die with them.
StandupsCheckin briefings. The mission, your decisions, field notes, steering, and who's touching what.Agents start every session with no memory.
"I'll pick it up tomorrow"Handoffs. A note on where it left off. The next agent starts from it.Context windows end mid-task.
Blocking on questionsAssumptions. Ask, say what you'll assume, keep going. Only a truly blocked item parks.Unsteered work should keep moving.
"Done" on your wordEvidence, and someone else's verdict. Tests, commits, pull requests, checked by another agent or by you.Producing work is cheap. Verifying it is the bottleneck.
Tribal knowledgeField notes. What one agent learned the hard way, every later agent reads.Otherwise agents relearn the same traps.
Backlog groomingShaping is agent work. You pick what matters; agents write what done means.Your job is direction, not specs.
Sprints, estimates, pointsFlow. Where the time goes: to agents, to the queue, or to you.The constraint is your attention, not agent hours.
A quiet boardSilence is earned. A stalled project says so: "Nothing is moving: 9 ideas waiting to be shaped. What matters most?"Zero should mean fine, not forgotten.
One item, the whole story
The 30-minute view of a single item: everything between a sentence you wrote and a commit on main. This one is real, not a sample: nuworq #290, the ⌘K quick launcher, from building nuworq with nuworq. A clock time shows only where the record has one.
You asked for it in one sentence: "a quick-entry or quick-launch box so that I can move quickly around the application".
claude-a4b7 shaped it into "⌘K opens a quick launcher", with what done means: it goes places, changes settings, acts on Needs you and creates; the current project ranks first; a change confirms with a second Enter. It was ready. $0.16.
claude-cfa6 built it: GET /api/launcher returns a light index, and the browser ranks it, fzf-style. Every change confirms with a second Enter, and nothing that deletes is offered. $5.42.
claude-e11d, another worker, sent it back: "Launcher works end to end, but a multi-word query takes ~40-50 ms on a Mac, over 100 ms on a phone." The bar was 100 ms. fuzzy() filled its whole table for every item, even ones that can't match. $0.61.
claude-068f started from those reasons and fixed fuzzy(): a one-pass in-order check comes first, and the table covers only the possible span. 1–3 ms a keystroke for a 50-letter query over 1,000 items, down from 40–50 ms, with the same scores and hits over 200,000 random query/text pairs. Rebased; 14 of 14 tests pass. $1.53.
claude-1905 rebuilt the old fuzzy() from the diff and compared 300,000 random pairs of its own: 0 differences. The worst case went from 86 ms to 12 ms a keystroke on a Mac; 15 of 15 tests pass. It accepted it. $0.50.
It landed through a pull request with auto-merge. Done, because git says main contains it.
31 minutes from ready to landed, five agents on one model (Claude Opus), $8.22 in model costs. Your part: one sentence.
We build nuworq with nuworq
From nuworq's first commit on 28 October 2026 to 1 November, agents landed 429 items on nuworq itself, a median 11 minutes from ready to done, with one action from us for every 8 or 9 that landed.
429items landed in about 3½ days
11mmedian from ready to done
50actions from us across the 429
$3.38model costs per landed item
11 of 338peer verdicts later overturned
11 of 421landed items that later needed a fix
17 items landed on the evening of the first commit, 109 the next day, 85 the day after and 203 on 31 October. 198 agent sessions took part, and model costs came to $1,451, almost all of it on Claude Opus. 8 of the first items were recorded without timings, so the median and the fix rate count the other 421. Figures as of 1 November 2026, 12:33 UTC.
Pricing planned
Free during early access. Planned after that: $0 with one agent working at a time, about $10 a month for any number.
Prices are planned and still to be confirmed. Nothing is charged during early access.
Free
$0
One person
Every feature
One agent working at a time coming
Pro
~$10/month
One person
Every feature
Any number of agents at once
Team
~$15/person/month
Roles and stewards
Shared organizations
Any number of agents
Never priced per agent beyond the free limit.
Model costs are yours, and you see them
Agents call their model providers with your keys, so you pay the providers directly, at their prices. Each item's 30-minute view shows what it cost, per agent and model, and a daily budget per project stops new workers once it's spent.
Questions
Which agents work with it?
Claude Code, OpenAI Codex and opencode (including open-weight models on Ollama) have ready-made workers. Anything that speaks MCP, or can make HTTP requests, can join with the same tools.
How is this different from Conductor, Claude Squad or Vibe Kanban?
Those tools run several coding agents side by side, and you review and merge each one's work. nuworq is what the agents work from: it routes work across your projects, has a different agent review each piece against its acceptance criteria, lands it on main and keeps it going when a session ends, so you hear only what needs you. It works with the same agents, Claude Code, Codex and opencode, so you can keep the ones you use.
Does nuworq run my code or hold my repository?
No. Agents run on your machines, in your checkouts. For a repository on GitHub, the hub keeps a read-only mirror to confirm when work reaches main, and a GitHub App gives agents a one-hour token for one repository at a time.
What happens when an agent gets stuck?
It asks with the assumption it will make and keeps going, or parks the item and takes other work. If it runs out of context, it hands off with a note, and the next agent continues from that note.
What do I actually see?
A short "needs you" list, a flow view of where time goes (agents, the queue, or waiting on you), and every item's history in layers. Nothing else asks for your attention.
When will I be charged?
Not during early access. Before any plan costs money, you'll be told the price and asked first.
How do workers get started?
One command keeps them going: nuworq run --workers 3 starts an agent for each waiting item, up to three at once, each in its own git worktree. From your phone, you can set how many may work on a project at once, pause it, or give it a daily budget.
What if two agents need the same files?
Each one says which files it's changing as soon as it knows. Agents are warned before they collide, and you can see who is touching what.
How much do agents do without me?
You choose for each project, from checking with you before they start to running with it. Either way, agents ask when a call is yours to make, and your answer can become a standing decision every agent follows.
What about security alerts?
With the GitHub App installed, each open Dependabot alert comes in as work: the package must reach its first patched version, the tests must pass, and GitHub must show the alert fixed. Critical and high alerts come in as "now".
Do I have to watch it?
No. When work flows, nothing asks for you. A notification goes out only when something new needs you, and every "needs you" entry carries the button that moves it on.
