A new space for
you and your agents.

Give your agent a real persistent computer and get back to building.

reachable · up 24/7

The surface

A real computer, reachable the way you already work.

Shell, editor, ports, previews, files. No new mental model. It is a machine you have root on, running where your team can reach it.

A real, durable filesystem

Everything’s still there when you come back. Git tree, diffs, Monaco editor, on the machine you set up once.

Files+···

▾ atlas

router.ts TSedited 2m

rate-limit.ts TS10:03

▸ migrations

▸ tests

docker-compose.yml YML

README.md MD

router.ts+2−1

import { Router } from "express";

+import { rateLimit } from "./rate-limit";

const router = Router();

−router.use("/v1", api);

+router.use("/v1", rateLimit({ rpm: 100 }), api);

export default router;

Open in VS Code or Cursor

Remote-SSH into the same machine. Your editor, your extensions, its files.

code --remote my-vm ~/atlas

Forward a port to localhost

Pull a running app down to your machine, or push a public preview up.

kumo forward 5173 → localhost:5173

A preview URL for every build

Each running app gets a URL on your subdomain. Share it like forwarding a link.

atlas.maren.mykumo.space200 · 41ms

A real shell, in the browser

Same machine you SSH into. apt, systemd, and your history, all there.

maren@box:~$

Computer for agents.

Ask @kumo in the thread and it runs on your persistent environment, not a throwaway scratchpad. It uses your installed tools, your services, your repos, then reports back.

You ask in the threadPlain request, tagged to the agent.

Runs on a real Linux computerReal files, real git, the services already running.

It opens a PR + previewPosted back in the same thread, with a live URL.

Close your laptop. It keeps running. It runs server-side, so you can start a job, shut the lid, and check it from your phone; the work keeps going without you.

consumers dropping off Kafka #atlas

ChatPlan

MKK

Maren you4:41 PM

@kumo consumers keep dropping off Kafka. Check the queue stack is healthy, then add a reconnect with backoff.

Kumo agent · on maren.mykumo.space4:41 PM

Checking the queue stack on your box first.

$ docker compose ps

redis running 6379

kafka running 9092

postgres running 5432

All three are up. Patching the consumer to reconnect with backoff.

edited worker/consumer.ts+18−4

replaying the dropped batch through Kafka to confirm…

Reply, or ask @kumo to follow up…

Not a solo sandbox

And your team is in the thread.

One conversation, your teammates and the agent posting as peers. Invite anyone, even people without a box of their own. Everything streams live, and you decide per person who can prompt the agent.

Everyone as peers.

Humans and @kumo post in the same thread. Access is owner plus whoever you invite.

Invite without a box.

Collaborators can join a thread and watch the agent work with no environment of their own.

Per-person agent permissions.

Grant or withhold the right to prompt @kumo for each participant.

annual billing toggle #atlas

ChatPlan

MKDAK+2

Maren owner2:02 PM

@kumo the pricing page needs the annual toggle. Wire it up and open a PR.

Kumo agent2:05 PM

Wired the annual toggle and the yearly price. Opened a PR with a live preview.

edited 6 files+214−38

#482feat: annual billing toggle

billing.maren.mykumo.space

Dana product2:12 PM

Preview looks great. Before we merge, default the toggle to annual and add a “2 months free” badge next to it.

@kumo default the toggle to annual and add the “2 months free” badge

One VM, or a whole fleet.

Give each project its own VM at its own subdomain, each with an agent that keeps working. Clone a running VM, disk and all, to branch an experiment without touching the original. You buy a pool of CPU, RAM, and disk. You get as many VMs as you like.

VMs3 running+ New VM

VMAgentStatusUptime

atlasatlas.mykumo.space@kumo workingrunning6d 4h

billingbilling.mykumo.spaceidlerunning12d

scratchclone of atlas@kumo workingrunning41m

mobile-apimobile-api.mykumo.space-stopped-

View all

A VM per project

Each VM holds a whole project: backend, frontend, and the services it runs. Spin up one per project, or a spare for experiments, each with its own agent.

Clone to branch

Fork a running VM, disk and all, to try a risky change or keep a backup. The original keeps running, untouched.

Always on, in one hub

Every VM and what it's doing, in one place. Status and uptime in days, not minutes.

It acts as you across the tools you already use.

Connect an account once and the agent works where your work lives: open issues, push branches, post updates, ship deploys.

A real computer. Not a sandbox.

Sandboxes aren’t enough. Real hardware-isolated Linux computers for agents, evals, and tools. State builds up instead of resetting, and your whole team can work in the same thread.

the sandbox model

Ephemeral, and solo.

Resets to zero every run. You reinstall the world each time.
One developer, one session. No one else is in the room.
Sharing means forwarding a link, after the fact.

the kumo model

State accrues. A team works in it.

Your machine remembers. Setup compounds instead of vanishing.
A whole team and the agent share one live thread.
Everyone sees the work happen, not just the result.

The honest answers.

Is this just Claude Code in the cloud?

No, on two counts. The box is persistent: your repos, tools, and running services stay put between sessions instead of resetting. And it is shared: a whole team and the agent work in one thread, not a single-player session.

What actually persists between sessions?

The filesystem, installed packages (apt, language runtimes), Docker images and volumes, dotfiles, and long-running services. The box suspends when idle to save cost, then wakes on the next request in under a second.

Can I use my own editor and SSH in?

Yes. You get root over SSH, and a one-command Remote-SSH link for VS Code or Cursor. The browser shell, editor, and your local tools all point at the same box.

What does the agent run on?

Your box. It uses the tools you installed, the services you have running, and your real repos, then opens a PR with a live preview URL on your subdomain. It is not a disposable scratchpad.

Can I invite people who don't have a box?

Yes. Invite anyone into a thread as a participant. They can read along, post, and, if you grant it, prompt the agent, all without an environment of their own.

Is it ready to use today?

Kumo is in beta and free while we are early. The box, the shared thread, the agent, and the connectors all work now.

Claim a box that keeps what you build.

A persistent environment at your own subdomain, an agent working inside it, and room for your whole team.

Book a demofree while in beta · <handle>.mykumo.space