Direct SSH access
You connect like to any machine. Your editor, your dotfiles, your usual flow.
sandbox · environment for agents
An SSH environment for your agents. Ready in seconds.
A persistent pod with SSH and its own volume that you manage. Start it empty or from a template (Claude Code, OpenCode, OpenClaw, Hermes) and it connects to your inference over an internal RDMA network, with an API key minted by the pod itself that never touches your laptop. It suspends itself when you're not using it.
getting started
From click to an SSH environment in seconds.
- 01
Pick resources (and a template, if you want)
In the wizard you pick resources (CPU, RAM, disk, GPU) and, if you want, an agent template (Claude Code, OpenCode…). You can also start an empty pod.
- 02
It provisions itself
It boots with SSH, DDN storage and an RDMA link to Token Factory. Pay per active hour; while it's suspended, it doesn't count.
- 03
Connect and work
Log in over SSH with your key and the agent is ready. When you're not using it, it auto-suspends.
what it is
A container you switch on, use, and that powers down on its own.
Codespaces pattern: you work over SSH, the Sandbox suspends after 30 min idle and comes back in seconds. Your persistent volume survives between sessions.
Persistent volume
Repos, datasets and checkpoints survive shutdown. Nothing to reconfigure when you return.
Auto-suspend
After 30 min idle it suspends and stops billing. Back in seconds when you reconnect.
templates
Pick a template and it boots configured.
Templates are an optional shortcut: instead of installing and wiring everything by hand, the Sandbox boots with the agent you choose already connected to Token Factory. If you'd rather, start from an empty pod and set it up your way.
Anthropic's coding agent with Claude models (egress to the Anthropic API).
Claude Code but on GLM 5.2 from the internal catalogue: no egress to Anthropic, all in the EU.
Open source coding agent, no lock-in to a single provider.
Agent environment for browser tasks and automation.
General-purpose agent template for custom workflows.
Need another template? Tell us and we'll add it to the catalogue.
connection with token factory
The Sandbox and the API, on the same network.
The Sandbox calls Token Factory over the cluster's internal RDMA network. Neither your API key nor your prompts hit the public internet.
No API key egress
The key travels over the cluster's internal network, never the open internet.
Minimal latency
API and Sandbox in the same datacenter in Spain, over an RDMA fabric.
One single credit
Sandbox and Token Factory draw from the same prepaid credit. One invoice.
storage
DDN EXAScaler for heavy data.
High-speed parallel storage, built for large repos, datasets and checkpoints without disk bottlenecks.
see it in action
Spin up a Sandbox in seconds.
Pick a template and the environment provisions itself: image, DDN storage and an RDMA link to Token Factory. Connect over SSH and get to work, your API key never leaves the perimeter.
gpuflow · sandbox-001live
$ gpuflow sandbox create --template claude-code
→ pulling image claude-code:latest
✓ image ready
→ attaching DDN EXAScaler volume
✓ /workspace mounted · 200 GB
→ linking RDMA to Token Factory
✓ RDMA up · API key never leaves the perimeter
✓ sandbox ready in 8.2s
$ ssh [email protected]
dev@sandbox-001:~$ claude --version
claude-code · connected to api.internal (RDMA)
pricing
You pay for active hours. Nothing else.
While the Sandbox is suspended, it doesn't count. Pay per active hour, billed at the end of the month. No upfront payment.
- From €0.15 per active hour (Starter)
- No upfront payment: add your card and pay only for usage
- Billed in euros, no commitment, no minimums
frequently asked
What people ask before spinning one up.
Which templates are there?
Claude Code, OpenCode, OpenClaw and Hermes, preconfigured against Token Factory. Need another one? Tell us and we'll add it to the catalogue.
Do I lose my work when it suspends?
No. The volume is persistent: repos, datasets and checkpoints survive shutdown. On reconnect everything comes back in seconds, with nothing to reconfigure.
How do I connect?
Over SSH with your public key, like any machine. Your editor, your dotfiles and your usual workflow.
Does the Sandbox have a dedicated GPU?
Not today. The Sandbox uses the GPU via Token Factory over the internal RDMA network. A dedicated GPU inside the Sandbox is on the roadmap, with no committed date.
How is it billed?
Per active hour, or a flat monthly plan with included hours. When it auto-suspends it stops counting. Same balance as Token Factory and dedicated GPU, in euros.
Where is my data?
In our datacenter in Spain, on DDN EXAScaler storage. It never leaves Europe.
Spin up your Sandbox in a couple of clicks.
Pick template, size and billing mode. SSH access reaches you by email.