Claude Code is built for your terminal — which is exactly the problem. Your machine has to stay on, stay connected, and stay attended. The moment you want an agent working while you're at dinner, on another project, or just asleep, you need Claude Code running somewhere that isn't your laptop.
There are three real ways to do that. Here's each one, honestly, including when you shouldn't use ours.
Option 1: Your own VPS
The DIY route. Rent a box from Hetzner, DigitalOcean, or AWS, SSH in, and set it up:
# On the VPS
npm install -g @anthropic-ai/claude-code
claude # log in with your Claude account, or set ANTHROPIC_API_KEY
From there, claude works exactly like it does locally, and claude -p "your task" runs headless. Pair it with tmux so sessions survive disconnects, and you have a persistent agent box you fully control.
Good: total control, cheapest raw compute, no third party anywhere in the loop.
Bad: you are now the platform team. Git credentials, branch hygiene, environment setup per project, reviewing whatever the agent did by reading diffs over SSH, and — the part everyone underestimates — the agent starts every run knowing nothing about your project except what you paste into the prompt.
Option 2: Hosted sandbox tools
Tools like Conductor give you managed cloud workspaces: isolated machines with your repo cloned, running Claude Code or other agent CLIs in parallel. You get a proper UI, workspace isolation, and PR-based review instead of raw SSH.
Good: zero infrastructure work, parallel agents, real review flow.
Bad: the agent still starts from the repo and a prompt. The plan — what you're building, what's already decided, what "done" means for this task — lives in your head or a doc the agent never sees. Vague prompt in, confident wrong answer out, at cloud scale.
Option 3: Cloud runs driven by the plan itself
This is the gap Scope Architect was built for. Instead of pointing an agent at a repo and typing a prompt, you assign it a task from your project's live scope — and the sandbox run starts with the full picture:
- Scope the project once. Modules, features, atomic tasks with acceptance criteria — generated with AI, refined by you in a workspace that reads like a project tool, not a wall of markdown.
- Link your GitHub repo and connect Claude Code or Codex in Settings → Cloud Agents.
- Dispatch a task. The sandbox spins up with your repo, your plan, your tech stack, your prior decisions, and that task's acceptance criteria.
- Watch it live. Status, logs, and comments stream into the Command Center. If the agent hits an ambiguity it flags the task and pauses — reply in the comments and the run resumes.
- Review the PR. Work lands on a dedicated branch and comes back as a pull request. Nothing touches your default branch until you merge.
The difference isn't the compute — a VM is a VM. The difference is what the agent knows when it starts, and where its questions go while it works.
Which one should you pick?
- You want a lab, and you enjoy infrastructure: VPS. It's a fine way to learn how these agents behave unattended.
- You want parallel raw agents on one repo today: a hosted sandbox tool does that well.
- You want the run to be right the first time — and visible while it happens: that requires the plan to travel with the task. That's the entire premise of Cloud Sandboxes, and it's included on every paid plan starting at $29/mo — runs bill to your own connected agent account or API keys, with no markup from us.
One warning that applies to all three options: unattended agents amplify whatever you give them. A sharp, scoped task comes back as a mergeable PR. A vague one comes back as 40 files of confident wrong. The cloud doesn't fix prompts — plans do.
Next: Claude Code keeps losing context? Here's the actual fix →