SandboxOverview

Sandbox

A real isolated computer for every agent. Each sandbox is a dev container or microVM with a shell, a filesystem, ports, snapshots, and optional GPU, pre-loaded with the coding harness of your choice. You drive it from an SDK, CLI, or dashboard. Sandbox is Tangle’s flagship product: the run in build, run, and improve.

What you get

  • A full machine per agent. Run shell commands, read and write files, open ports, snapshot state, and attach a GPU only for the command that needs it.
  • Any coding harness. Each sandbox runs one AI backend: OpenCode (default), Claude Code, Codex, Cursor, and a dozen more. You pick it per sandbox.
  • Durable agent sessions. Dispatch a prompt and reconnect to the same run after a client crash, a deploy, or a browser reload. Same session ID is idempotent, so a retried request never re-executes.
  • One key, metered per second. The same sk-tan- key authenticates every call; you pay for what you use.

Why use it

  • Give an agent a real environment, not a chat box. It runs actual commands in an isolated machine and returns real results: files changed, tests run, output captured.
  • Run untrusted work safely. Every sandbox is isolated with its own filesystem and egress policy, so an agent’s mistakes stay contained.
  • Keep long runs alive. Durable sessions survive the things that normally kill a streaming agent: a dropped connection, a redeployed worker, a closed laptop.
  • See what happened. Every run can emit a trace to Intelligence, so a failed run is inspectable instead of lost.

How it works

Your SDK or CLI call hits the Sandbox API, which routes to a driver (container or microVM), provisions the machine, and relays the agent’s output back to you as it streams. You never manage the infrastructure: create returns a machine, delete releases it. For the execution and orchestration internals, see the runtime docs.

Use cases

  • Coding agents that ship real changes. Point Claude Code, Codex, or OpenCode at a repo in a sandbox and let it edit, run, and verify.
  • Batch and background work. Dispatch many runs across a fleet of sandboxes, each isolated, each metered.
  • Evals at scale. Run a task set across sandboxes and send every trace to Intelligence to score what changed.
  • GPU jobs on demand. Attach a spend-capped GPU lease around the one step that needs it, then release it.

Start

  • Quickstart. Create a sandbox and run your first agent in a few minutes.
  • SDK reference. The core client, sandbox, session, and GPU methods.
Quickstart