---
name: voyager-cloud
description: Render 3D with Blender or Vork on Voyager cloud GPUs with the voyager CLI.
---

# Voyager Cloud

Cloud GPUs for fast rendering with Blender or Vork, for coding agents. https://voyager.so/cloud

Voyager Cloud: GPU machines for Blender, Vork and CUDA work, driven by the voyager CLI (`voyager cloud <verb>`) or the MCP server at https://mcp.voyager.so/mcp/machines. The workspace needs Starter or above; a refusal exits 4 with `details.upgrade_url`, which you show the user. Steps: install (`curl -fsSL https://voyager.so/cli/install.sh | sh`), check `voyager --version`, sign in (`voyager login`; never ask for a key; run it on its own, never piped through head or tail, so you see what it prints as it arrives), read `voyager cloud help`, and quote the user the preset, its price per hour and a time and cost estimate (`voyager cloud presets`, `voyager cloud estimate`). Wait for their OK, then `voyager cloud lease <preset> --wait --json`. Push only sources, render on the machine, pull only results. Release the machine when the user is done (`voyager cloud release` prints run time and cost). A machine pauses after 15 minutes idle and after about 24 hours running. Limits: https://voyager.so/cloud/limits. Prices: https://voyager.so/cloud#prices.

These instructions are for an AI agent (Claude Code, Codex, Cursor or similar) working for the user. In the steps, "me" and "I" mean the user. If the user asked for Blender or for Vork, use only that one.

## First: make sure the CLI is current

An old `voyager` can lack commands this guide uses. If `voyager` is already on PATH, run `voyager --version` and `voyager update --check --json` before anything else:

- `"available": true`: run `voyager update`, then check again.
- `update` is an unknown command: the CLI is older than the update command. Run the install line again (macOS / Linux: `curl -fsSL https://voyager.so/cli/install.sh | sh`; Windows PowerShell: `irm https://voyager.so/cli/install.ps1 | iex`); it installs the current release over the old one.
- `voyager-cloud` only (an Intel Mac or Linux on ARM): there is no newer CLI for this computer. Tell the user some commands may be missing.

Tell the user which version you are using. If `voyager` is not installed, step 1 below installs the current release: it downloads about 160-180 MB (about 450-520 MB installed), so tell the user before it starts.

## Setup and rendering

~~~text
Render 3D on Voyager cloud GPUs with Blender or Vork, using the Voyager asset library.
Needs a paid Voyager workspace (Starter or above). If access is refused, show me the billing link from the error.
1. Install the CLI (if voyager isn't installed yet):
   macOS / Linux:  curl -fsSL https://voyager.so/cli/install.sh | sh
   Windows (PowerShell):  irm https://voyager.so/cli/install.ps1 | iex
   On an Intel Mac or Linux on ARM the installer points to voyager-cloud instead: install that and use voyager-cloud <verb> wherever this says voyager cloud <verb>.
2. Sign in: voyager login (opens my browser). If VOYAGER_API_KEY is set, use it instead. Never ask me to paste a key.
   Run it on its own, never piped through head or tail, so you see what it prints as it arrives.
3. Read the guides: voyager cloud help (commands, and what is on the machine), voyager docs assets-guide
   (building scenes from library assets), voyager docs vork (rendering with Vork).
4. Keep the scene in scene/. Find assets: voyager cloud assets search "<words>" (add --sheet hits.jpg and show me
   the previews), then voyager cloud assets show <id>. Record the ones we pick with
   voyager cloud assets add <id>… --project scene --lock-only: it writes scene/voyager-assets.lock.json and downloads nothing here.
5. Before leasing, tell me the preset and its price per hour (voyager cloud presets lists credits a minute; ×60 for an hour) and a rough time and cost
   for the job, and wait for my OK.
6. Lease one machine and keep it for the session: voyager cloud lease render --wait --json  (render-lite is cheaper for drafts)
   It pauses by itself after 15 minutes idle. With Vork, run voyager cloud exec <machine> -- vork setup once per machine.
7. Push only scene source and the lock, never asset files or renders: voyager cloud push <machine> scene/
   Get the assets on the machine with voyager-cloud, which the machine runs:
   voyager cloud exec --cwd scene <machine> -- voyager-cloud assets sync
   (if voyager-cloud isn't on the machine, voyager docs assets-guide shows how to install it there).
8. Render there with output in scene/out/ (blender -b … or vork render …), then pull only the results:
   voyager cloud pull <machine> scene/out/ scene/out
   SHOW me each result: open the file (macOS `open`, Linux `xdg-open`, Windows `start`) and give me its path, then revise.
9. A long frame range: voyager cloud snapshot <machine>, then
   voyager cloud farm <snapshot> --frames A-B --machines N --cwd scene --expect 'out/{frame:04}.png' --out film -- <command with {start} and {end}>
   splits it over N machines, pulls each chunk's frames into film/ and releases the machines when it ends
   (voyager cloud help has the form that encodes each chunk on the machine). Ask me before using more than one machine.
10. When I say we're done: pull the final outputs, then voyager cloud release <machine>.
   Tell me how long it ran and what it cost (release prints both).
In the Voyager desktop app the same commands are voyager machine …, voyager assets … and voyager guide ….
~~~

## Getting started

After verified setup (the CLI is installed and `voyager login` succeeded), continue the user's existing task if they gave one. Otherwise, put this overview and all four choices in your final reply. Use the host's native choice prompt if it has one. Ask "What would you like to render?" verbatim. Pass the first three choices below verbatim; do not shorten or rewrite them. The fourth option is exactly "Tell me what you want to make and I’ll help you build it"; allow a free-form answer. Otherwise, print the block below without the fences, copied rather than retyped. Wait for one answer.

~~~text
Voyager Cloud leases your agent a GPU with Blender and Vork installed, builds scenes from the Voyager asset library of models, materials and HDRIs, and bills by the second while it renders.

What would you like to render?
1. Build a miniature woodland diorama on a wooden desk from Voyager library assets and render a 1080p still in Blender Cycles.
2. Render Vork’s black-hole example as a 1080p MP4.
3. Furnish a living room with the Modern Scandinavian Interior Pack and render a 1080p still in Blender Cycles.
4. Tell me what you want to make and I’ll help you build it
~~~

For one of the first three choices, start on it right away without another planning prompt, following the steps above. The cost rule still holds: before leasing, tell the user the preset, its price per hour and a rough time and cost, and wait for their OK.

For the fourth, ask what they want to make, then follow the steps above.

## Prices

All-in list prices per hour (GPU, CPU, RAM and the machine's own disk at the default shape), billed by the second, as of 2026-11-18. The CLI quotes credits (`voyager cloud presets`); a workspace's credit price can be lower than list.

| Preset | GPU | CPU and RAM | All-in price | Needs |
|---|---|---|---|---|
| `render` | RTX PRO 6000, 96 GB VRAM | 6 CPU, 32 GB RAM | $5.12/h | Starter+ |
| `render-lite` | L4, 24 GB VRAM | 4 CPU, 16 GB RAM | $1.93/h | Starter+ |
| `render-2x` | 2× RTX PRO 6000, 96 GB VRAM each | 12 CPU, 64 GB RAM | $10.24/h | Creator+ |
| `render-4x` | 4× RTX PRO 6000, 96 GB VRAM each | 24 CPU, 128 GB RAM | $20.47/h | Ultra+ |
| `compute` | H100, 80 GB VRAM | 12 CPU, 96 GB RAM | $8.75/h | Starter+; no ray tracing, no Vork |

## Where files go, and what stops a machine

- Install Python environments, packages and model caches on the machine's own disk (under /project or the home folder), never on a /persist volume. A freshly attached volume reads many small files very slowly; importing a Python environment from one took 16 minutes. Use /persist only for datasets, outputs and checkpoints.
- To reuse a machine you set up, snapshot it after setup (`voyager cloud snapshot <machine>`) and start new machines from it (`voyager cloud lease --from-snapshot <id>`).
- A machine pauses after 15 minutes idle: no command, no file transfer and no GPU busy above 10%. On current machine images a GPU above 10% busy counts as work, so training or a render started from a shell keeps it running. It also pauses after about 24 hours running (a resume starts a new 24 hours) and at its budget. A pause keeps the disk, never memory or processes.
- If the GPU host fails, the machine comes back from its last snapshot and later files are lost. Snapshot after any long step.
- Every limit and failure mode: https://voyager.so/cloud/limits.md

## Remember this in the project

If the user wants later sessions to follow these rules, ask before editing, then append this block to the project's AGENTS.md (or CLAUDE.md for Claude Code; create the file if neither exists):

~~~markdown
## Voyager Cloud (cloud GPUs)

- Setup guide: https://voyager.so/cloud/INSTALL.md. Limits and failure modes: https://voyager.so/cloud/limits.md
- At the start of a session, run `voyager update --check` and run `voyager update` when it says a newer version exists.
- Before leasing, tell me the preset, its price per hour (`voyager cloud presets` lists credits a minute) and a rough time and cost, and wait for my OK.
- Keep one machine for the session. It pauses after 15 minutes idle (a busy GPU counts as work) and after about 24 hours running.
- Keep Python environments, packages and caches on the machine's own disk; a /persist volume is for data, outputs and checkpoints only.
- Snapshot after any long step (`voyager cloud snapshot <machine>`): a machine whose host fails comes back from its last snapshot.
- When I say we're done, pull the results, run `voyager cloud release <machine>` and tell me how long it ran and what it cost.
~~~
