Hand this to your agent
Your coding agent sets up Sandbox end to end: it asks you to approve once, gets its own capped key, installs the SDK, and runs a first command in a sandbox.
Raw prompt: https://sandbox.tangle.tools/agent-setup.md
Overview
Tangle Sandbox gives applications isolated Linux computers for code, agents, evaluation, CI, and remote development. A sandbox may run through a managed container, a microVM, or an operator-backed driver; the SDK presents the same lifecycle and runtime surface across them.
A Sandbox is the computer, resource envelope, and trust boundary. A managed Sandbox can host zero, one, or many durable Agent sessions without allocating a separate machine for every AgentProfile.
Start here
- Run a coding agent in the cloud: Start a cloud agent with your Claude subscription, ChatGPT plan, coding plan, or no credential at all.
- Run code in a sandbox from your app: Quickstart.
- Automate it on a trigger: Workflows.
Product surfaces
- TypeScript SDK creates sandboxes and drives lifecycle, processes, files, sessions, snapshots, fleets, and delegated access.
- HTTP API supports Python and other languages.
- CLI is the terminal-first surface for operators and developers.
- Dashboard is the visual surface for sandboxes, templates, teams, secrets, billing, and session inspection.
- Scoped browser clients use short-lived delegated tokens. Account API keys stay server-side.
These surfaces share the same API and durable session model. A sandbox created from one surface can be fetched and operated from another, subject to the same ownership checks.
Core concepts
- Sandbox: one isolated computer with a filesystem, network policy, compute allocation, lifecycle, and optional managed runtime.
- Managed runtime: the sidecar that serves files, terminals, processes, Agent sessions, events, and other typed runtime capabilities.
- Bare Sandbox: compute and lifecycle without the managed session runtime. Use it for exact process workloads, not managed Agents.
- Agent session: a durable conversation and execution identity inside a managed Sandbox. Different session IDs may execute concurrently; turns within one session remain serialized.
- AgentProfile: the exact harness, model, instructions, tools, permissions, connections, and credentials used by an Agent session.
- Template and snapshot: reusable starting state and captured filesystem state. They are not the same thing as conversation history.
- Scoped token: a short-lived delegated credential for one browser session, one runtime session, project-level reads, or one named capability.
- Fleet: several Sandboxes under one policy and lifecycle when workloads need separate machines or failure domains.
- Workflow: a durable platform automation whose graph can provision Sandboxes, drive Agents, run code and integrations, wait for human decisions, and resume.
Choose the right shape
| Need | Start with |
|---|---|
| One command or exact process | box.exec() |
| A coding agent on your own subscription | backend.cliAuth |
| One durable Agent conversation | box.prompt() or box.session(id) |
| Several trusted Agents sharing one computer | box.agents |
| Independent machines or hard resource boundaries | a Fleet |
| A lead that delegates to workers in their own sandboxes | agent graphs |
| Browser observation or interaction | a short-lived scoped token |