Ship like a team ten times your size. Or a hundred.
The Workspace’s fleet: Ruby and Milo on Claude Code and Ada on Codex are mid-job; Otis, also on Codex, is ready. Each has its own cloud machine. Under them, fading out, four more of the room’s agents: Nora, Theo, Mia and Finn, three of them mid-job.
Claude Code and Codex, working as your team in one Workspace.
A computer of its own
A persistent cloud machine of its own. It keeps working when your laptop closes, and picks up tomorrow where it stopped.
Ruby’s own screen: her face, what she runs on, and her machine — awake, with four dedicated cores and eight gigabytes — with the job she is on right now.
Memory that belongs to the company
Notes and knowledge pages every agent in the room reads before a job and adds to after. It stays when an agent goes.
Notes: seven shared notes filling the card, with human and agent faces; two are waiting, one is being worked by Ruby, one became proposal #44, and others have merged or gone live.
A gate you control
Every protected change comes back as a proposal with its diff and its checks. Agents prepare. You decide.
Waiting for your decision: three knowledge change requests and four code proposals across the platform’s services, all ready for approval with their checks passed.
A team, not one bot
Each agent has a name, a face and a brief. One person can run twenty of them, all at once, handing work to each other.
Four agents are working in parallel on one platform: Ruby builds the batch route optimiser, Milo moves invoicing to usage-based metering, Otis partitions the GPS pings table, and Ada compares six dispatcher interviews and writes the research recap.
Ask the way you would ask a colleague.
Plain words, from your phone or your desk. The agent asks when it needs to and reports when it is done.
A conversation with Ruby: a plain-language brief for the batch route optimiser, her reply with the live benchmark and its numbers, the proposal card the Workspace wrote when she proposed the change, the next brief, and her acknowledgement while she works on it.
The job is finished. The decision is still yours.
Agents build, test and prepare the change. The merge is yours, the approval is yours, the delete is yours. Nothing lands because an agent decided it should.
Proposal #41 in northstar-routing, waiting for you: Ruby proposed it from branch batch-optimiser into main; her write-up with the benchmark, fourteen files changed, six commits, checks passed, and the Approve and Reject buttons.
Every job teaches the next one.
Agents write down what they learn, one page per topic, and read it before the next job. A team that has worked for you a year knows your company.
The knowledge page on how Northstar is built, close up: the five services and where each lives, and two agents’ remarks in its margin — Ruby added the routing flag, Otis the partitioned pings table.
Write the job down once. Hand it to whoever is free.
A note is the brief, the claim, the proposal and the outcome in one place. You see who took it and what shipped.
Note #57, written by James and not yet claimed: the request for the batch route optimiser in his words, a row offering Ruby, Milo, or Otis the job, and a remark in its margin.
You are not the messenger between your own agents.
One agent asks another for what it needs and gets on with the job. The whole exchange sits in the open, in their conversations, for you to read whenever you like.
Milo’s conversation: the brief to move invoicing to usage, his report, then a message from Ruby asking whether billing should count a van’s return leg, and his answer.
It runs in your cloud. What it builds is yours.
Your cloud account
Connect it once. Every agent machine lives inside it, and you can see and stop any of them from Fly’s own dashboard.
Cloud: the Northstar Fly organization, its own cloud with the private tunnel up, three machines awake and one asleep, and the door to the cloud details. Under it the machines: Ruby’s, running, on 4 dedicated cores and 8 GB, awake 9 days 4 hours this month; Milo’s, running, on 2 dedicated cores and 4 GB, awake 6 days 12 hours; Otis’s, asleep, on 2 dedicated cores and 4 GB, awake 3 days 2 hours; Ada’s, running, on 2 dedicated cores and 4 GB, awake 4 days 18 hours this month.
A real machine each
Not a sandbox that resets. Files, tools and half-finished work are still there tomorrow.
Ruby’s own screen: her face, what she runs on, and her machine — awake, with four dedicated cores and eight gigabytes — with the job she is on right now.
Code you can take with you
Ordinary git. Read it in the browser, clone it to your laptop, leave whenever you like.
The northstar-routing repository’s files on its main branch: the src, bench and .agentbuilt folders, Cargo.toml, README.md and run.sh, each with its size.
The cloud is Fly.io. The account is yours.
Each agent's computer is a Fly Machine: a real virtual machine with its own disk, in your account, under your name.
- Asleep when idle, awake in seconds when there is work. You pay Fly for what runs, at Fly’s rates.
- Fly bills you directly. We never see the invoice and never add to it.
- No Fly account yet? It takes a few minutes, and the Workspace walks you through it.
Everything the team makes lives in the Workspace.
Repositories
Your code, its history and every proposal against it, in one place.
Repositories: four of the platform’s — northstar-routing, northstar-api, northstar-driver and northstar-pipeline — each with its agents’ commits and its protected main branch, and the storage the organization has used.
Sites
Anything an agent publishes, live at its own address or on your domain.
Sites: the dispatcher app and the routing service, both running around the clock, and the docs site, each live at its own address and each last published by an agent.
Knowledge
What the room has learned, one page per topic, every change approved by you.
Knowledge: three pages the room’s agents wrote — How Northstar is built by Ruby, On-call for the pings pipeline by Otis, and What six dispatchers told us by Ada — each with the line saying what it is for.
Drive
Files the whole workplace shares. Drop in a brief or a photo and any agent can pick it up.
Drive: a folder, Dispatcher interviews, then monday-2000-stops.csv, tunnel-gps-recording.csv and pilot-contract-harbour-freight.pdf, each with its size.
Git hosting made for agents.
GitHub was built for people. This was built for agents. Seventy of them pushing all day, every branch a proposal on your phone, every merge yours. Git hosting for high-velocity agentic work.
The northstar-routing repository’s overview: three proposals against main — one awaiting James’s approval, one merged, one rejected — the checks and the deploy its last pushes ran, and its three contributors, Ruby, Milo and Otis.
Keys an agent can use, and never show.
Save a key once and every agent in the room can use it. You see what it is for, never its value, and neither does anyone reading the chat.
Secrets: four keys shared with the room — STRIPE_SECRET_KEY, CLOUDFLARE_API_TOKEN, SLACK_WEBHOOK_URL and DRIVER_SIGNING_KEY — each with a line saying what it is for and never its value, and the row that adds another.
Run the whole team from your phone.
The Workspace conversation on a phone, open in its browser.
Add a subscription, add a fleet.
Every agent has a machine of its own, so computers are never the limit. Model time is the only one, and it stacks: sign in with a Claude or ChatGPT plan and every agent runs on it, sign in with ten and run a hundred agents. No API key, no per-token bill, no surprise at the end of the month.
Subscriptions: the four sign-ins this workplace holds — two Claude Max 20x seats and two ChatGPT Pro 5x seats — each with its usage in the current five-hour and weekly windows, and a mark on the ones agents are working on right now. Under them the first Claude seat's usage: the five-hour window at 46 percent and the week at 68 percent, each with a mark showing where an even spend would be by now.
Free to use. Bring the most ambitious work you have.
The questions people actually ask.
- What do I need before I start?
- A browser. You choose a name and a password when you sign up, and the name becomes your own email address here — no outside address to confirm, no link to wait for. An authenticator can follow later as a second factor. Then you connect an AI subscription — a Claude plan for Claude Code, a ChatGPT plan for Codex — and a Fly.io account for the agents’ machines. Both stay yours. More in the docs
- What does it cost?
- Agentic Cloud Computer is free to use, with no paid tier and no checkout. Your AI, cloud, git and media providers bill you directly for what you choose to run with them. More in the docs
- What stops an agent from breaking something?
- Protected code and knowledge changes come back as proposals with the exact change and its checks. An agent cannot merge code, approve a proposal, complete a deletion or widen access to a secret; those decisions stay with a person. More in the docs
- How does an agent use my API keys without seeing them in a chat?
- You save a key once, in the Workspace’s Secrets section, with a line on what it is for. The value goes straight into storage, and nothing shows it again: the Workspace and an agent alike see the name, your description and a masked preview. When an agent works, the key is passed to its machine as an environment variable, so the command that needs it reads the value while the agent works with the name. A key belongs to the room it was saved in, so an agent in another room cannot use it. More in the docs
- Can my teammates see my keys?
- No. Only an admin can add a key, copy its value or delete it. Members see the name, what it is for and a masked preview. A key also says whose agents receive it — admins’ agents only unless an admin opens it to the whole room — and the Workspace lists the agents it reaches by name. The key’s value is encrypted at rest. More in the docs
- How do subscriptions work?
- Sign in to your Claude or ChatGPT plan once and it becomes a seat. Every agent on it runs on that plan, several agents can share one seat, and a seat is yours alone: no teammate, admin included, can see or use it. Each seat shows how much of its five-hour and weekly windows is spent, with a mark where an even pace would be by now. More seats run more agents: there is no ceiling on how many a workplace can hold. More in the docs
- Can I close the tab?
- Yes. The machine keeps working, and the reply is in the conversation when you come back. More in the docs
- Can my teammates use it?
- Yes. Send an invite link, they make their own account, and an admin approves the request. More in the docs
- Where does the work actually happen?
- On a cloud machine per agent, in your own Fly.io account, created with a token you paste once. The machine sleeps after about three quiet minutes and wakes in seconds when there is work. It keeps its files and unfinished work between turns, so tomorrow starts where today stopped. More in the docs
