Get started
Docs
Sandscale is isolated cloud compute for AI agents. Start a sandbox for the code your agents write, give coding agents a persistent cloud computer, and keep every run governed and audited.
QuickstartSign up, install the CLI, start a sandbox, run a command and connect your coding agent.
CLI referenceEvery
sandscale command, generated from the CLI itself.
Using the APIHosts, authentication, errors, limits and retries for api.sandscale.dev.
API referenceEvery endpoint by area, with parameters, fields and error codes, generated from the server’s code.
GuidesRun code in a sandbox and lease a build host, step by step with the API.
MCP toolsThe tools sandscale mcp offers your agent.
What you can run
| Product | What it is |
|---|---|
| Sandboxes | Firecracker microVMs for the code your agents write. A separate kernel per run. |
| Workspaces | Persistent cloud computers for coding agents and people. Files survive; open them in the browser or over SSH. |
| Build hosts | Clean, leased build machines for agent tasks, released when the task ends. |
| Lab hosts | Your own Linux or macOS machines, with their hardware, claimed per run. Team and Enterprise. |
| Agent Cloud | Give coding tasks to agents in your workspaces and get a branch and pull request back. Pro and up. |
| Object storage | S3-compatible storage for files, datasets and build artifacts that outlive a run. United States now; Europe coming soon. |
Ways in
| Way in | Where |
|---|---|
| Console | cloud.sandscale.dev |
| CLI | sandscale for macOS, Linux and Windows. See the Quickstart. |
| MCP | sandscale mcp connects Claude Code, Codex, Cursor and other MCP tools. |
| API | api.sandscale.dev, with an API key or your own sign-in. |
| SDKs | TypeScript (@sandscale/sdk) and Python (sandscale): coming soon. |
Safe by default
- The network is off by default for sandboxes. On Pro and up you choose what an agent may reach with an egress allowlist.
- Every run gets its own kernel in a Firecracker microVM, not a shared container.
- Secrets live in Vault: referenced, injected at run time, never placed in prompts or logs.
- DLP on prompts and tool calls (Enterprise).
- Approval gates and policies on templates, sizes, run times and which agents may run.
- A hash-chained audit log of every action, which you can export as evidence.
- A kill switch for your organization.
- Your data stays in the region you choose.
Sandscale is not open to the public yet. These docs describe the first public release. To get in early, email sales@sandscale.dev.