Rubeau
Testers
App screenshot: the company’s building in 3D, at night, seen from above, with its lit rooms and its staff.

The office, rearranged. Your AI staff at work, right in front of you.

  • Windows
  • in testing, by invitation
  • built on Claude
  • your own API key or a local model

Real app capture — the building at night

The tour

Watch Rubeau run

Two and a half minutes inside the app, narrated by a synthetic voice, with captions.

Join the testers

Creating the company, the team, a real past mission replayed from its log, the letter to the boss, the report, rules and memory, day and night, layout mode, the voices, access and data.

  • Replay: no mission was started for this video; the app replays the logs of real past missions (recorded in French).
  • Demo profile: a copy with its personal data removed.
  • Filmed on 23 September 2026, on version 0.401.73, before the three stages of 28 September: what the tour calls “the beta” is now the tester phase, free and by invitation, with a thirty-day licence.

Synthetic voice: Pocket TTS, by Kyutai, licensed under CC BY 4.0 (opens in a new tab). Full screen makes the interface text readable.

The idea

A company of AI agents you run without writing code: describe what you want, a chief takes the mission, gathers colleagues, delegates, and you get the result — knowing who did the work, who is stuck, and what it used.

Not a graph of agents. A company you watch at work.

Every agent is a permanent employee with a name, a face, seniority, a mission history and a running total of usage. You see the same faces every time you open the app. Your agents are not disposable: they have careers.

The building

An office that is always there

Nothing that moves on screen is decoration: it shows what is really happening in the company.

  1. The whole company, on its street

    The same rooms and the same people every time you open it. The company keeps its staff from one mission to the next.

    Real app capture · mission data recorded in French

    App screenshot: the company’s building in 3D among its streets, roof open, during the replay of a mission, with the plan, the messages between colleagues and the log on the right.
  2. A finite number of desks

    Every room has its capacity, shown next to its name: places taken out of the total.

    Real app capture · mission data recorded in French

    App screenshot, close-up: the building’s rooms in 3D with their labels — Leadership 2/4, Engineering 1/6, Lounge 3/10, Board room 0/12 — and the staff at their desks.
  3. Day and night

    At night, only occupied rooms stay lit.

    Real app capture · mission data recorded in French

    App screenshot: the building at 10:30 PM, during the replay of a mission; only occupied rooms are lit.
  4. The building is yours

    In layout mode: furniture, walls, doors, rooms, and the floors too. Nothing is placed until you save.

    Real app capture · mission data recorded in French

    App screenshot: layout mode and its toolbar — turn, remove, add furniture, draw a room, place a door, save or discard.

Real app capture · mission data recorded in French

App screenshot: the company’s building in 3D among its streets, roof open, during the replay of a mission, with the plan, the messages between colleagues and the log on the right.App screenshot, close-up: the building’s rooms in 3D with their labels — Leadership 2/4, Engineering 1/6, Lounge 3/10, Board room 0/12 — and the staff at their desks.App screenshot: the building at 10:30 PM, during the replay of a mission; only occupied rooms are lit.App screenshot: layout mode and its toolbar — turn, remove, add furniture, draw a room, place a door, save or discard.

Between missions

  • A coffee break, a game of ping-pong: these show people who are genuinely idle, never a random animation. Ping-pong takes two idle people in the same room.
  • The company can have working hours. If you set them, everyone goes home at night and the floor empties: no new mission starts, unless you override the hours.
  • Closing is not killing: a mission in progress runs to the end, even after hours.

No hours are imposed by default: it’s your company, so set them if you wish.

A mission

How a mission unfolds

You say what you want. You see who will lead the mission, and the amount at which work gets cut off. Then the team organises itself right in front of you.

  1. The director plans, then delegates

    She writes a plan, then hands each part to a colleague, chosen from their file. What each one is doing shows above their head.

    Real app capture · replays of real past missions, recorded in French

  2. The director writes to you

    One question, up to four replies to choose from; the mission waits for your answer, costing nothing.

    Real app capture · replays of real past missions, recorded in French

  3. The report, with its sources

    At the end, the report comes back to the CEO, with its sources.

    Real app capture · replays of real past missions, recorded in French

Answer a letter

The letter to the CEO

A decision is yours to make

A chief is instructed not to ask your permission to work. They write only when the decision is truly yours: two incompatible directions, a scope so ambiguous that guessing would waste the budget, a finding that contradicts the mission, or a choice that is hard to undo.

One letter, one question: the one only you can decide.

Mock-up — fictional example

Inès Marchal is calling you

From
Inès Marchal, Director — mission chief
To
The CEO
Subject
Mission "comparison of 3 CRMs for a small business"
Used so far (API equiv.)
$0.37 (sample value)
Deadline
she decides on her own in 9 min

Two of the three CRMs don’t publish their prices for ten users. Do I compare all three on their public plans, or widen to CRMs that publish everything?

Why I can’t decide on my own Either choice changes the deliverable, and guessing what you want would waste the budget.

In the app, a decision letter comes with a free-text reply field; folder and delivery requests are answered with buttons only. Nothing is sent from here: this page calls no model.

See who did the work

And if nobody replies?

No reply from you — the deadline has passed

The chief decides alone, taking the most cautious, most reversible option, and her final reply states which question went unanswered and what she assumed.

The rules of a letter

  1. One question, written for someone who hasn’t followed the work.
  2. Up to 4 genuinely different replies: you answer in one click.
  3. At most 3 calls to the CEO per mission: beyond that, the chief decides alone and says so.
  4. While it waits, the mission is paused and consumes nothing.
  5. No reply within 10 minutes? The chief decides alone and writes down what they assumed.

In the app, a chief can also read their call aloud, in a voice synthesised on your machine, offline.

Sample letter — mock-up. Its rules, buttons and deadline come from the app, translated from French; the header layout, names and amount are fictional.

Real app capture

App screenshot: “Ada is calling you”, a letter from the director to the boss, with her question, four proposed replies, a free reply and “Decide for yourself”. The letter itself stays in French: it is a recorded fact.
The real letter, in the app — replay of a past mission (the letter itself stays in French: it is a recorded fact)

The staff

The file decides

The personnel file decides who gets the work — not chance, not whoever arrived last.

The AI model belongs to the person and is chosen at hiring: the same company employs different profiles side by side.

You hire, you review, you make the call. The product informs; it never forbids.

Mock-up — fictional example

  • Chief

    Inès Marchal

    Role
    Director
    Model (chosen at hiring)
    top tier
    Verdict
    reliable
    Seniority
    18 days
    Missions
    14 missions
    Failure rate
    8% failure
    Tool calls
    96 tool calls
    Total used (API equiv.)
    $11.20
    Per mission
    $0.80/mission

    13 missions concluded, 8% failure, $0.80 per mission (averaged over 14 missions).

  • Théo Garnier

    Role
    Researcher
    Model (chosen at hiring)
    economy
    Verdict
    improving
    Seniority
    18 days
    Missions
    22 missions
    Failure rate
    18% failure
    Tool calls
    171 tool calls
    Total used (API equiv.)
    $2.64
    Per mission
    $0.12/mission
    Trend
    trend: 36% → 0% failure

    improving: 36% failure before, 0% since.

  • Maëlle Roux

    Role
    Legal counsel
    Model (chosen at hiring)
    mid tier
    Verdict
    unproven
    Seniority
    3 days
    Missions
    2 missions
    Failure rate
    0% failure over 2
    Tool calls
    14 tool calls
    Total used (API equiv.)
    $0.96
    Per mission
    $0.48/mission

    2 missions concluded — too few to judge (3 are needed).

Real app capture

App screenshot: the staff files — hiring, promotions, and for each employee their role, model, seniority, tool calls, failure rate and cost.
The real staff files, in the app (the promotion lines are still in French: a known beta translation gap)

A file’s verdicts

unproven
not enough concluded missions to judge
reliable
delivers, rarely fails, and costs little per mission
improving
fails markedly less than before
expensive
delivers, but costs a lot per mission
uneven
delivers, but fails too often to be called reliable
declining
fails markedly more than before
failing
fails half the time or more

A mission you stop is not the fault of whoever led it: it counts as neither a success nor a failure.

Mock-up — fictional data; the fields and verdicts are the app’s.

Safeguards

Nothing leaves without you

Three clauses, enforced by the code, not just promised.

  1. The team writes only in the folder you grant it

    When a chief needs to write files, they ask you in a letter. You choose the folder yourself, or let them carry on without writing.

    I need a folder to write files in. Will you grant me one?

    • Choose a folder…
    • Carry on without writing

    the app’s buttons, not clickable here

    No reply, no folder: a right is never won by waiting.

  2. Budgets are capped

    Every mission has its own budget, set at launch. The company also has a monthly cap: once it is reached, no mission can start.

    $0.37

    Used $0.37Stops above $2.00

    illustrative values, in API-equivalent dollars, as in the app

    The budget is checked between work turns: a turn already under way runs to completion, so the total can go over the budget.

  3. Nothing is delivered before you have seen it

    When the work is done, the chief shows you what was written to your folder — file names and sizes — and asks whether to hand it over. What the chief says and what Rubeau has checked are shown separately.

    The work is done. Here is what was written to your folder — shall I hand it over?

    What the chief says — their words, unverified:

    “The comparison is ready, with an appendix on pricing.”

    What Rubeau has checked:

    • crm-comparison.md — 6 KB
    • pricing-appendix.md — 3 KB

    sample files

    CEO’s tray — The document is waiting for your decision.

    No reply means the work is not handed over.

Mock-up: the letters’ questions and buttons are the app’s; the amounts (in API-equivalent dollars: with your own key, an estimate of what the provider bills you; with a local model, nothing) and the files are examples.

The log

Everything is an event in the log

The same source feeds the building on screen, the personnel files and the usage.

Mock-up — fictional example

  1. Time Who Inès MarchalEvent mission startedDetail comparison of 3 CRMsFeeds: Building · Files
  2. Time Who Inès MarchalEvent takes the roleFeeds: Building · Files
  3. Time Who Inès MarchalEvent work planDetail 3 stepsFeeds: Building
  4. Time Who Inès MarchalEvent hands the work onDetail to Théo GarnierFeeds: Building
  5. Time Who Théo GarnierEvent takes the roleDetail Open spaceFeeds: Building · Files
  6. Time Who Théo GarnierEvent uses a toolDetail web searchFeeds: Files
  7. Time Who Théo GarnierEvent usageDetail $0.04Feeds: Files · Usage
  8. Time Who Théo GarnierEvent has finishedDetail report returnedFeeds: Building · Files
  9. Time Who Inès MarchalEvent asks you to decideDetail one question, 3 repliesFeeds: Building
  10. Time Who CEOEvent you answeredDetail reply no. 2Feeds: Building
  11. Time Who MissionEvent resumesDetail after waiting 1 min 32 sFeeds: Building
  12. Time Who Inès MarchalEvent mission finishedDetail $0.91 in totalFeeds: Building · Files · Usage
  • Building10movements on screen
  • Files7personnel-file entries
  • Usage2usage lines

The log is append-only: nothing is erased, nothing is rewritten.

Mock-up — lines reconstructed for this page: the event labels and the time format are those of the app’s log; names and amounts are fictional. Amounts are in API-equivalent dollars, as in the app: with your own API key, an estimate of what the provider bills you; with a local model, nothing.

Real app capture

App screenshot: the live log in the right-hand column, while colleagues search the web; one line per event, with its time.
The real log, in the app — replay of a past mission (recorded in French)

Today

Where Rubeau stands

Project status

4,600+ automated tests

  • In testing, by invitation
  • No public release yet
  • Sales not open yet
  • Windows app
  • Built on Claude
  • Your own API key or a local model
  • Demo on request
Current milestone
The tester phase, by invitation, with a thirty-day licence. Next: the beta, then the stable release, when sales will open.

Timeline

Show the earlier milestones (8)
  1. First commit: a multi-agent engine with no interface, under a code name.

  2. The office on screen, and a first real mission launched from the app.

  3. The company takes shape: the personnel file decides who works, the office has working hours, and the chief writes to the CEO.

  4. The isometric city, the chiefs’ voices, and a new name: Steading.

  5. The building takes shape in 3D, and people sit down at their desks.

  6. The building gets its glass lobby, and people walk in 3D instead of teleporting.

  7. Core decisions: the team writes only in the folder you grant. And the delivery letter: nothing is handed over until the CEO has seen it.

  8. More than 2,000 automated tests, and Steading becomes Rubeau.

  1. Access rules clarified with Anthropic: a distributed copy runs with your own API key or a local model, never with a subscription.

  2. Isolation probe passed: a colleague sees only the task it is given, and the guard stops it outside the project folder, for reading and writing alike.

  3. First project taken all the way through: Rubeau’s team translates into English, and reviews, the guide for its own testers.

  4. The narrated tour on this page is filmed inside the app.

  5. A single Windows installer, voices included, and three stages: the testers, the beta, then the stable release.

The founder’s own letter — not a mock-up

A note from the founder

I’m Javad Razavi, and I have been building Rubeau on my own since June 2026.

I wanted to run AI agents like a team you know well: the same faces, a personnel file, a cost, and a letter when a decision is mine to make.

What exists today: a Windows app that runs real missions on Claude, a company that keeps its staff, and a log where everything is written down.

What doesn’t exist yet: a public release and sales. First the testers, by invitation; then the beta; then the stable release.

The videos and screenshots on this page come from the app; the interactive scenes are mock-ups, and each one says so: I would rather show honest work in progress than a product built on promises.

If you want to try it, challenge it or simply talk about it, write to me. I read and answer every message myself.

Javad RazaviFounder of Rubeau

For investors

Rubeau is for people who don’t write code but have work to hand over. Once an app runs a whole company of agents, the real question becomes: who do I keep, who costs too much, who keeps failing? A developer can read logs to find out; everyone else can’t. Rubeau answers that question with a measured personnel file, finite capacity and a letter to the CEO.

Today, the app runs on Windows, carries out real missions on Claude and rests on more than 4,600 automated tests. There is no public release yet, and the project is funded by its founder.

I’d like to talk with people interested in joining or backing the project. The address is the same for everyone: a first conversation is simply a chance to get to know each other and look at the product together.

Get in touch

You are…

You’ll get a message when a test build is ready. For now, it’s Windows only.

You will need your own API key (your usage is billed to you by the model provider; before each mission, you set the amount at which work is cut off) or a local model on your machine: free, but the chief then works alone, with no tools and no colleagues.

Tell me what you’d like to hand over to a team: that’s what helps me most.

The subject line of your email will be pre-filled based on this choice.

Gallery

The screenshots, full size

All the real app screenshots, to enlarge. The work scenes are on-screen replays of real past missions.

  • Overview3D overview — app screenshot · mission recorded in French
  • The 3D view3D view — app screenshot · mission recorded in French
  • A letter to the CEOLetter to the CEO — app screenshot · mission recorded in French
  • The staff filesStaff files — app screenshot · promotion lines still in French
  • The logMission log — app screenshot · mission recorded in French
  • At nightThe building at 10:30 PM — app screenshot
  • Layout modeLayout mode — app screenshot
  • A floorLevel 1 — app screenshot
App screenshot: the company’s building in 3D among its streets, roof open, during the replay of a mission, with the plan, the messages between colleagues and the log on the right.
Overview3D overview — app screenshot · mission recorded in French1 of 8

Contact

Get in touch

No form, no third-party mailing list: your message goes from your own email straight to me. I read everything and reply personally.

Subject: Rubeau — I’d like to test it

Are you an investor? Same address. Write as an investor

Or copy the address:contact@rubeau.fr

Want to see the app running? Ask for a demo