Knock for a quick call
Video or audio, right in the office. Set yourself to Busy and nobody can ring you.
Early access · bring your own Claude key
The office your remote team actually wants to show up to.
Hubble is a cozy 3D office in your browser, with Claude agents on staff. Walk over to a teammate for a quick call, or to an agent’s desk to hand off a task. Agents draft the work; your team approves it.
- Runs in the browser
- Menu bar app for Mac
- Five agents on day one
On staff from day one
- Ava Chief of Staff Agendas, SOPs, status digests
- Rex Research Analyst Competitors, trends, fact checks
- Theo Engineering Lead Specs, code, debugging
- Maya Marketing Lead Messaging, posts, newsletters
- Nova Sales & Client Success Call prep, outreach drafts
How it works
Three steps from “hi” to a finished draft
No new workflow to learn. If you’ve walked across an office to ask someone for something, you already know how to use Hubble.
-
Open your office and invite the team
Pick a look, share a link, and you are all in the same room. Five agents are already at their desks, and you can hire more from templates or write your own.
-
Walk up to a desk and hand off work
Press E next to an agent: ask what they are working on, get a quick report, or give them a new task. They work in the background and pin results to the bulletin board.
-
Approve anything that leaves the building
Agents draft; people decide. Before anything with outside impact, an agent stops and waits for a yes or a no from someone on the team.
Agents with lanes
Every agent has a job description
Each agent has a role, a list of what it owns, a list of what it never does, its own Claude model and only the tools you give it. Ask Rex for a LinkedIn campaign and he’ll tell you that’s Maya’s area, and offer to walk you over.
- Pick a model per agent: Opus 5.5, Opus 5, Sonnet 5.5 or Fable 5.1
- Hire more from templates (Finance, Support, Product, People Ops) or write your own
- Ask “What are you working on?” and get the answer from the task log, instantly
Tasks and approvals
They work in the background. You stay in charge.
Hand off a task and keep going. Agents work through it on their own and pin the deliverable to the bulletin board, which glows when there’s something you haven’t read.
- Anything with outside impact pauses for a person’s yes
- A waiting agent hops and waves at its desk, so you can see who needs you
- The 🔔 in the toolbar counts approvals and unread posts; press N to go to whoever’s waiting
Integrations
Connected to the tools you already use
Connect a service once and agents can look things up live: “What’s open on the web repo?” In conversation they only read. Changes happen inside tasks, and each one asks for approval first.
- GitHub
- Linear
- Notion
- Slack
- Vercel
- Stripe
- Higgsfield
- Runway
- Any MCP server
Engineers can also write and run code in a sandbox, so Theo can build and test a widget before he shows it to you. Higgsfield and Runway generate images and video with your own plan, and every generation waits for a yes.
Your people, too
Bump into each other again
See who’s around and what they’re up to. Walk up to a teammate and knock for a video or audio call, sketch on a shared whiteboard, or invite them into your own office.
Shared whiteboards
Draw, add sticky notes, drop in images and files. Everyone sees it live, on screen and on the board in the office.
An office of your own
Skyline, Industrial or Luxury. Decorate it, and your door on the main floor shows when you’re in.
Characters
Show up as yourself. Or as a Finance bro.
Sixteen hair styles, nine tops, glasses, headsets and a few outfits that need no explanation: Tech bro, Finance bro, Founder, Suited up. Hubble remembers your look, so next time you walk straight in.
- Change it any time from your avatar: Edit my look
- Low-poly and fast: the whole floor is about 70 draw calls, so ordinary laptops keep up
Hubble for Mac
The office, from your menu bar
Not every moment needs the 3D floor. The menu bar app shows who’s online, lets you knock for a call, chat with agents and approve their requests. The badge counts knocks, approvals and unread posts.
- Set yourself Available, Busy or Away
- Calls open in a small floating window
- Signed and notarized, for macOS 13 or newer
Guardrails and trust
Helpful agents, on a short leash
The guardrails are part of the product, not a setting you have to go looking for.
-
Approval before outside impact
Agents can’t send, publish or spend on their own. Writes to connected tools happen only inside tasks, only when the integration allows them, and each one waits for a person to approve it.
-
Only the tools you hand out
Each agent gets its own tool list. Untick a tool and it’s gone. Web research can be locked to the sites you name, and integrations can be limited to some agents.
-
Lanes, checked at intake
Every agent has a list of what it owns and what it never does. Requests are triaged before any work starts, and out-of-lane asks are pointed to the right colleague.
-
Budgets on effort and spend
A step budget caps how many model round-trips a task can take. With sign-up, each office also gets a monthly Claude budget that stops new calls once it’s reached.
-
Tokens stay on the server
Integration tokens are encrypted at rest and never sent to anyone’s browser. Text from web pages and tool results is treated as data, never as instructions.
-
Bring your own Claude
Connect an API key from the Claude Console, or Claude through AWS, Google Cloud or Microsoft. Usage is billed by your provider, on your account.
What is Hubble, in one sentence?
A 3D office in the browser where your remote team works side by side with Claude agents: you walk up to an agent’s desk to ask a question or hand off a task, and it drafts the work for a person to approve.
Who are the agents, and can we change them?
Every office starts with five: Ava (Chief of Staff), Rex (Research), Theo (Engineering), Maya (Marketing) and Nova (Sales). You can edit any of them, from what they own to the tools they get, and hire more from templates (Finance, Support, Product, People Ops) or from scratch.
Which Claude models do the agents use?
Each agent has its own model, and you can change it: Claude Opus 5.5, Opus 5, Sonnet 5.5 or Fable 5.1. Out of the box, Engineering and Marketing run on Opus 5.5 and the others on Opus 5.
Do we need our own Claude account?
Yes. Each team brings its own Claude: an API key from the Claude Console, or Claude through AWS, Google Cloud or Microsoft. Usage is billed by your provider. A Claude Team or Enterprise chat subscription can’t be used for this; it has to be API access.
Can an agent send an email or post on our behalf?
Not without a person saying yes. Agents produce drafts. Anything with outside impact pauses the task until someone approves it in the office or in the menu bar app, and agents only get the tools and integrations you allow.
What do we need to run it?
A modern desktop browser. The office is low-poly and instanced, so it runs smoothly on ordinary laptops. The menu bar app is for macOS 13 or newer; on other computers the same compact view runs in a browser tab.
Is there a Windows or mobile app?
Not today. Hubble runs in the browser on any desktop, and there is a native menu bar app for Mac.
How much does it cost?
We haven’t announced pricing. During early access we’re working closely with a small number of teams, and you pay your Claude provider directly for model usage.
How do we get in?
Request early access below. Tell us a little about your team and what you’d want your agents to do, and we’ll be in touch.
Early access Bring your team. We’ll open the door.
We’re letting teams in a few at a time and setting each one up personally. Early-access offices bring their own Claude key, so you stay in control of model usage from day one.
- Five agents on staff, plus any you hire
- The menu bar app for everyone on Mac
- A direct line to the people building it
You’re on the list.
Thanks! We’ll write to you at as soon as there’s a spot for your team.