Skip to content

Guides

Computers

Markdown

A computer is a machine of its own for a coding agent. Create one, connect Claude Code to it, work in it over a shell or its desktop, and let it shut down when idle; the home disk stays.

Some work needs a place whose files are still there tomorrow: a checkout, a half-finished migration, the notes of a long task. A computer is that place. It is an app with one machine that runs the computer image, a home disk mounted at /data that outlives the machine, and a desktop you can watch or take over from the console.

Size vCPU Memory Home disk
small 1 2304 MB 50 GB
base (default) 2 4096 MB 100 GB
large 4 8192 MB 128 GB

An organization on starting credit may run 2 computers at a time; see Limits and quotas.

Open Computers, choose New computer, give it a name and a size. The console creates the app, the home disk and the key, and deploys the computer image.

Names are 2 to 31 characters of lowercase letters, digits and hyphens. The home disk is a volume named <computer>-home; it is kept while the computer is shut down and when its machine is replaced.

What the image brings: Node.js with npm, Codex, OpenCode, git, the GitHub CLI, Python 3 with pip, a browser and a desktop. Claude Code is not pre-installed; inside the computer, rytiv-setup claude installs it with its own official installer under its own terms (--stable, --version <v> and --print are accepted), and rytiv-setup alone shows what the image has and has not.

Two ways, for two kinds of work.

  1. Claude Code on your laptop, Rytiv as its tool. Add the MCP server once; the agent then creates computers, runs commands in them and snapshots them with the tools computer_create, computer_exec and computer_snapshot:

    claude mcp add --transport http rytiv https://api.rytiv.app/v1/mcp --header "Authorization: Bearer <your key>"

    Then ask in plain words: “Create a computer on Rytiv, clone the repository into it and run the tests.”

  2. Claude Code inside the computer. Open a shell in it and install Claude Code there; the session keeps running when your laptop is closed:

    ry ssh dev-1

    Inside: rytiv-setup claude, then claude. Sign in with your own Claude account; nothing of it touches Rytiv.

A computer the MCP tool computer_create makes is an agent machine: it does not sleep on its own, and it can take snapshots. A computer from the console or ry computer create shuts down when idle and keeps its home; it takes no snapshots.

ry exec dev-1 -- git status
ry exec dev-1 --dir /data -- npm test
ry ssh dev-1
ry forward dev-1 3000:3000

ry exec runs one command and exits with its exit code; ry ssh opens an interactive shell over the Rytiv API, no SSH keys needed; ry forward brings a port of the computer to your machine without a public port. Each of them wakes a shut-down computer. Through MCP, computer_exec does what ry exec does, as an argument list without a shell.

Files you want to keep go under /data, the home. The root disk is the image’s and starts fresh at every boot.

ry computer view dev-1
ry computer control dev-1
ry computer release dev-1
ry computer audit dev-1

view gives a link to watch, for any number of viewers, valid 60 to 3600 seconds (default 900). control gives a link to take keyboard and mouse: there is one holder per machine, the agent inside pauses while you have control, and --force takes control from the current holder. audit shows who got which link and how control ended. The console shows the same desktop on the computer’s page.

After 10 minutes without activity the computer shuts down; your files stay. Every size is above the 2048 MB up to which a machine sleeps with its memory, so a computer stops instead: open programs end, and the next ry exec, ry ssh, request or console visit boots it fresh with the home disk as it was. CPU load, a held control lease and an open viewer with frames flowing count as activity.

While it is shut down it costs nothing for memory and CPU; the home disk is billed as a volume. See Credit and prices.

A machine created with durableCheckpoint: true (what the MCP tool computer_create does, and POST /v1/apps/:name/machines with that field) can be snapshotted and restored:

Method Path What it does
POST /v1/machines/:id/checkpoints {label, memory, comment}. Takes a snapshot now.
GET /v1/machines/:id/checkpoints The machine’s named snapshots
POST /v1/machines/:id/checkpoints/:label/restore Brings the machine back to that snapshot. With keepCurrent: true (the default) the current state is saved first.
DELETE /v1/machines/:id/checkpoints/:label Forgets a snapshot

Disk only (memory: false, the default) is a short pause, and a restore boots fresh with the files as they were. With memory (memory: true) a restore resumes the very processes; the pause grows with the machine’s memory. Labels use letters, digits, ., - and _, start with a letter or digit, and are at most 63 characters; auto- is reserved. Through MCP, computer_snapshot takes one. Such a machine also survives a host failure with its last automatic checkpoint.

ry apps destroy dev-1
ry volume destroy vol-1a2b3c4d

Deleting the app removes the machine; the home volume stays until you delete it too (ry volume list shows it). Both cannot be undone; the CLI asks you to type the name, automation passes --yes.