AI agents your technical teams already use, run as a company asset

The work keeps going.Everyone can see it.

AI, one laptop at a time: each agent lives in one person's window, stops when that laptop sleeps, and nobody else sees what it did. Oasis makes those agents a shared way of working: they keep going on machines you own, and the whole team can watch, read and steer them.

The agents here are the ones engineers and technical staff run on a computer to write code, run tasks and fix things: Claude Code (from Anthropic), Codex (from OpenAI) and Kimi Code. Your company keeps paying those providers directly; Oasis is the shared workplace they run in.

Companies license Oasis per rollout; anyone can try it on their own machine first.

Download for your own machine →

macOS, Linux & Windows

Nothing to sign up for to use the app. Installs on your Mac, Linux or Windows PC; works with Claude Code, Codex and Kimi Code.

Live in your browser

This is the real thing.
Click around.

The actual Oasis app in demo mode, with a sample company fleet: four machines and several agents. Nothing to install, nothing to sign up for. Press Play, then open any window. Open the full-page demo to click around.

oasisgui.com/demo Open full ↗

Loading the live demo…

The Oasis workspace: the company's machines listed down the left, four jobs tiled in the middle, and one AI agent reporting that its tests passed.
Open full-page ↗

Every screen on this page is captured from this built-in demo fleet; the colleagues' cursors are staged to show presence.

What changes

What changes
when the AI is shared.

The easy way to adopt AI is the way spreadsheets were adopted: one person, one window, one laptop. Good for the individual, invisible to the company.

Today

  • Each agent lives in one person's chat window and stops when that laptop closes.
  • Nobody else can see what it did, what it was told, or what it changed.
  • Work stops when the laptop shuts, the Wi-Fi drops, or the person goes on leave.
  • Each person's agent runs under a personal account with its own settings, which the company cannot see or keep consistent.
  • A new colleague starts from zero; a leaving one takes the context with them.

With Oasis

  • Agents run on company machines and keep going while people are away.
  • After a restart they come back where they were, with the same conversation.
  • Everyone invited sees the same screen, live, and reads the work as plain prose.
  • Every machine, every agent and every person, in one window.
  • Access belongs to people: given for a set period, so it lapses unless renewed, and withdrawn in one place.

The same agents your teams use now: Claude Code, Codex, Kimi Code. Oasis launches them, brings them back and reads them; it does not resell them. The AI providers you subscribe to still process what your agents send them: Oasis changes where the agents run and who can see their work, not where the models run.

Continuity

Close the laptop.
The work carries on.

Each agent runs on one of your machines, not inside the app on someone's laptop, so closing the app changes nothing. Disconnect, sleep, go home: the agent keeps working. After a restart, agents and their conversations come back where they were, with the settings they started with.

The Oasis window: the company's machines listed down the left and one AI agent's window filling the rest, ending with its report that the work is done and 74 tests passed.
The company's machines down the left; one agent's window on the right, just finished and reporting back. Close the app and this keeps running on that machine. From the built-in demo fleet.
  1. 01 Brief Give an agent a task, or give an AI manager a goal and a schedule.
  2. 02 Agents work On company machines, or in a throwaway virtual computer that is wiped afterwards. They keep going when the laptop closes.
  3. 03 People steer Read the work as prose, reply, try a second direction, hand it to a colleague.
  4. 04 You are told One notification centre says when an agent finished or is waiting for you.

Give it a goal. Check the board.

An AI manager (we call it a Supervisor) keeps a goal and a schedule, hands tasks to agents, and tracks them on a board of to do, in progress and done. It keeps working while nobody is logged in, using an AI model you choose, on an account you hold.

After a restart, it comes back.

When a machine reboots, its sessions return with the same names, folders and conversations. An agent picks up where it left off, with the settings it started with.

Try two directions at once.

Copy an agent's conversation at any point and run two versions side by side, a safer one and a bolder one. Both keep everything said so far.

Told when you are needed.

One notification centre collects "the agent is waiting for you", finished tasks, builds and machine events, each linked to its machine and session. Every desk sees it.

A record you can search.

Every machine keeps a searchable library of its agent conversations, with recording switched on by you. Find every place a topic came up, across the company.

A sandbox when you want one.

Give an unattended agent its own clean, throwaway computer on a company server, with extra computing power for heavy jobs when needed; on a server set up for it, one starts in seconds. Its mistakes stay off your other machines, and you delete it when the job is done.

A Supervisor session named Release Captain: its goal and next wake-up time at the top, and a board with To do, In progress and Done columns, one card marked blocked.
An AI manager with its own board: to do, in progress, done, one item blocked, next check in a few minutes. It wakes on a schedule and reports back, using an AI model you choose.
An agent's conversation shown as prose: the request in a bubble, the agent's numbered plan, a folded group of technical steps, and a reply box at the bottom.
What the agent did, as a conversation. The request in plain language, the plan, the result, and a box to reply that goes straight to the running agent. Anyone can read this.
The notification list open on its All tab: an agent finished, a build completed, a context-limit warning, a deploy and a machine coming back online, each labelled with its machine and session.
You are told when an agent finished or is waiting for you, with the machine and session named. Every desk sees it.

If a machine restarts, the conversations and sessions come back; anything an agent had not yet saved at that instant is lost, as with any program. A machine that is switched off is off. When you reconnect you see the recent part of the conversation on screen; the full record is in the searchable library.

Oasis does not stop an agent from acting: it shows you everything the agent does but adds no extra permission step of its own. The limits you set inside the agent still apply. For work nobody is watching, give the agent its own throwaway computer so a mistake stays off your other machines.

Shared work

Same screen.
Different chairs.

A machine's workspace is one shared picture that everyone invited sees as it changes: the agents, the files, the notes, the board. Your team looks over an agent's shoulder together, from any desk, with no screen-share call.

The top of the workspace with two colleagues' avatars in the toolbar and their named cursors resting on an AI agent's window and a build window.
Alice and Milan are here now, looking at the same agent you are. Their cursors are live. Nothing to set up, nothing to send.

Shared desktops.

Invite colleagues to a shared desktop on a company server and they see the same windows, files and board, live. Each colleague gets their own named login on that server, so access is per person and can be taken away per person; there is no share link to forward.

People, not machine names.

Faces in the toolbar and live cursors on the work. Colleagues appear as people, not as machine names.

Hand work over.

A colleague takes over mid-task from another desk, ticks off the shared to-do list, or hands the next item to the agent in one click.

A whiteboard with the work on it.

Notes, checklists and risks next to the agents' own windows, drawn and typed together as it happens.

Private team chat.

Chat and direct messages travel over your own private network and are encrypted between people; the server that passes them along cannot read them.

Onboard in one step.

One invite code gives a colleague their login, their identity and their place on the private network; access is withdrawn from the People panel, in one place.

A shared whiteboard with sticky notes, a ship checklist and a risk note, the agents' live windows framed beside them, and two colleagues' cursors.
A shared whiteboard: a ship checklist, a risk note and the agents' own windows pinned beside them.
A two-pane file browser on one of the company's machines with two colleagues' cursors over the same listing.
Browsing a machine's files together: same folder, same scroll, two cursors on one listing.

Shared desktops run on a Linux server you administer; Macs and Windows PCs join them. Treat invite codes like passwords.

A scenario built only from what ships. Not a customer story.

One ordinary day,
four chairs.

Nothing below needs a vendor account, a vendor cloud or a new operations team.

  1. 01
    Morning. The owner, at home.

    The sidebar shows every machine and who is on each. The overnight AI manager finished two tasks; one agent is waiting for an answer. She reads its plan as prose and replies: yes, but keep the old pricing table.

  2. 02
    Mid-morning. The team lead takes over.

    Same machine, different chair. He sees her cursor, ticks a shared to-do, hands the next one to the agent, and copies the conversation to try a bolder approach beside the safe one.

  3. 03
    Afternoon. The new hire joins by one code.

    One invite code creates her login, her identity and her place on the private network. The whiteboard already holds the notes and the live agents; she draws on it from a Windows laptop.

  4. 04
    Night. The server restarts.

    An update reboots the machine. By morning the sessions are back with their names, folders and conversations, and the team carries on from the board.

Ownership

Your machines.
Your network.

Oasis reaches your machines with the secure connection your IT team already uses (SSH) and joins them on a private network you run yourself: office, home and cloud. The computer that admits machines to that network is one of yours, not ours. No vendor cloud sits between you and your machines.

A map of nine company machines with lines showing which can reach which, and badges on each marking its role and whether the internet can see it.
A live map of the company's machines: which are reachable, which the internet can see, and who routes through whom.

Access belongs to people.

Roles for owner, admin, operator and member follow the person, not the laptop. Access to a machine is given to a named person for a set period and taken away in one place. A person's identity can be kept on a small USB security key.

See what the internet can see.

Each Mac or Linux machine can be locked to the private network only, and a scan from your own server shows what the internet can reach. Windows PCs rely on Windows' own firewall.

Backups you control.

Each machine backs itself up on a schedule, encrypted, to storage you choose and hold the keys to. Restore one file or a whole session from any backup point.

Usage you can see.

Claude Code, Codex and Kimi Code stay your subscriptions. One view shows how much of each one's allowance is used on each machine and which have room right now. Oasis sells no AI of its own.

Settings that stay in line.

Keep a company default for Claude Code and Codex settings on every machine. Where a machine drifts from it, you see the difference; an unauthorised API-key override is reset in a click.

The secure connection IT already uses.

Oasis reaches machines over the keys IT already manages (SSH). Switch Oasis off and the machines, sessions and files are still there.

The Backups settings page with one storage destination configured and three machines each switched on with a scope and a daily schedule.
Three machines backing themselves up daily, encrypted, to storage the company chose.
The People and computers panel: three machines listed with their roles of owner, admin and operator, an Invite a computer button, and a legend explaining the four roles.
Decide who may do what: owner, admin, operator, member. Adding a computer is one single-use code, not a server login.
The Accounts dropdown listing the company's Claude, Codex and Kimi subscriptions, each with usage bars and the machines it is signed in on.
The company's Claude, Codex and Kimi subscriptions in one place: which have room right now, which are near a limit, on which machines.

What leaves your machines. In plain terms: your code, conversations and files stay on your machines. Oasis (made by Varrock Inc.) hears from the app only when it checks for updates. Exactly, with the two small exceptions: the desktop app has no user accounts, no telemetry and no analytics. The desktop app's only contact with Varrock-operated infrastructure is the update check against oasisgui.com; everything else runs on machines you control — except that the Mesh view asks api.country.is for the country of each peer's public IP to draw a flag, and Claude accounts you add are checked against Anthropic. Other connections happen only when you switch a feature on: the AI model you pick for an AI manager, your backup storage, public image servers when you create a virtual machine, and public-address checks for the exposure scan. The security documentation lists every one.

Getting started

Start with one person.
Finish with the company.

Oasis installs like software, not like a service: no account, no vendor cloud, no server to buy before the first install. Your IT lead installs the app, adds the company's machines, and Oasis puts a small background program on each one over the connection IT already uses; nothing is added to a machine without IT choosing it. Grow into it in four steps.

The Invite a collaborator dialogue with the name alice typed in, listing the account, key and network invite one code will create.
One code for a new colleague: their login on the machine, their identity and their place on the private network.
  1. 01
    One person, their own machines.

    Install on a Mac or Linux laptop and add the servers IT already connects to. From then on, agents run on those servers and outlive the laptop.

  2. 02
    One shared machine.

    Pick a Linux server and invite the team with single-use codes. Shared desktops and the whiteboard switch on.

  3. 03
    The private network.

    Create it in Settings and run its control point on a server you own; office, home and cloud join. Add a second control machine for safety. Lock machines to the network.

  4. 04
    The fleet.

    Backups to your storage, every AI seat in one view, documentation for IT. Talk to us about licensing and a rollout.

Mac: Apple Silicon. Linux: x86-64. Windows PCs join and view; the work itself runs on Linux or Mac machines. Updates are announced and installed only when you choose.

For your company

What we have.
What we don't yet.

Your IT and security reviewers will want the detail: how it is built, what talks to what, how access is granted and taken away, how to run it day to day. Varrock Inc., the company behind Oasis, writes that documentation against what actually ships and rates each security measure as present, partial or missing. Request access for your company; your IT lead signs in once approved.

A settings page titled All my terminals listing twelve sessions on four of the company's computers, grouped by computer, with a status beside each one: working, idle or your turn.
Every session on every company machine in one list, grouped by computer, each with its agent's status.

What ships today

  • Access by named person, for a set period, taken away in one place.
  • A private network you run yourself; each Mac or Linux machine can be locked to it.
  • Encrypted backups to storage you choose, with file and session restore.
  • A searchable record of agent conversations on each machine, with recording switched on by you.
  • A Mac build signed by Varrock and notarized by Apple; an updater that verifies every download.
  • A desktop app with no user accounts, no telemetry and no analytics.

Not yet

  • No single sign-on. Your people cannot log in with their company Microsoft or Google account; each person gets their own key, optionally on a small USB security key. Joining and leaving are done by hand, not from your staff directory.
  • No compliance attestations. No SOC 2 or ISO report. The security reasoning is written down in the documentation. Legal documents are drafts for review and say so.
  • Windows PCs view and steer; agents themselves run on Linux or Mac machines. Windows shows a warning the first time because the installer is not yet signed with a purchased certificate.
  • No export of activity logs to your security tools, and no packages for IT's device-management systems; logs stay on each machine and in the app.

We'd rather you knew before you asked.

Rolling it out to a team? Write to sales@oasisgui.com with the size of your team and the machines you run; we reply with licensing and a rollout plan. Your reviewers can request access to the documentation.

sales@oasisgui.com

Requesting access asks for your name, work email and company; see the privacy policy.

Free for personal use

Try it on
your own machine.

Free for personal use on macOS, Linux and Windows. For teams and companies we offer enterprise licensing: talk to us. Licence terms. Updates are announced, verified, and install only after you choose Update & Restart.

macOS Apple Silicon

Signed with Varrock's Apple Developer ID and notarized by Apple. Local virtual machines need macOS 13 or newer. Intel Macs are not supported.

Linux x86-64

Needs libwebkit2gtk-4.1, libgtk-3, libayatana-appindicator3, libssl3, openssh-client. NixOS: use the flake; updates then follow Nix.

Windows x64

Installs for the current user. Windows joins and views; the work runs on your Linux or Mac machines. Windows will ask: More info → Run anyway.

macOS, Linux & Windows

macOS: Apple Silicon · Linux: x86-64 · Windows: x64

Teams and companies: talk to us about enterprise licensing.

Built in Rust

A native app. The engine, the private network and the helper on each machine are written in Rust. On a Mac the terminal can draw with the native graphics engine.