Sandbox API

Overview

One HTTP API for benchmarks, agents, code execution, and long-lived sandboxes (pods and VMs).

Sandbox runs code and AI agents in isolated environments and returns the results. You send a request; the platform provisions a substrate (Kubernetes pod, cloud VM, or external Mac), runs the work, scores it when there is a verifier, and keeps artifacts for you to download.

Everything is one REST API behind a single hostname per environment:

export SANDBOX_GATEWAY_URL="https://sandbox.turing.com"   # or your environment's host

What you can do

I want to…CallGuide
Run a benchmark task and score itPOST /sandbox/harbor/v2/jobs/execute-tasksRun a benchmark
Run a coding agent (no verifier)POST /sandbox/agent/v1/runsAgent runs
See which agents / profiles / languages existGET /sandbox/v1/catalog?kind=…Supported agents
Run a snippet of code oncePOST /sandbox/codeedit/v1/executionsCode execution
Keep a workspace across many commandsPOST /sandbox/sandboxes/v1Long-lived sandboxes
Open a GUI desktop (Linux pod / Windows VM / Mac)POST /sandbox/sandboxes/v1 + desktop-* profileDesktop sandboxes

Where work actually runs

The API is the same; the substrate depends on the profile / driver hint:

SubstrateTypical useExamples
Kubernetes podHarbor trials, Agent runs, CodeEdit, Linux desktop, default long-liveddriver_hint: k8s
Cloud VMWindows desktop (RDP via Guacamole)driver_hint: ec2-vm (also gce-vm / vm)
External Mac cloudmacOS desktop streamdriver_hint: external + provider aws-ec2-mac
Daytona / dockerOptional Harbor environment overriderequest field sandbox: daytona

Details and diagrams: Architecture.

Supported agents (catalog)

Agent runs accept only curated catalog ids. Harbor can also accept Harbor factory agents when enforcement is warn. Always discover from the live API:

curl -sS -H "X-Api-Key: $SANDBOX_API_KEY" \
  "$SANDBOX_GATEWAY_URL/sandbox/v1/catalog?kind=agent"
Catalog idDisplay nameNeeds model?Notes
opencodeOpenCodeYesRecommended default for ad-hoc runs
claude-codeClaude CodeYesCLI pinned in catalog
codexOpenAI CodexYesReasoning models only for default effort
gemini-cliGemini CLIYes
openhandsOpenHandsYes
cursor-cliCursor CLIYes
aiderAiderYes
cline-cliCline CLIYesCertified with W&B Inference
terminus-2Terminus 2YesHarbor reference agent (LiteLLM)
oracleOracleNoNeeds task solution/solve.sh — validates tasks, not free-text

Full table, enforcement rules, and examples: Supported agents.

Start here

  1. Get started — first request, end to end
  2. Architecture — components, pods, VMs, flows
  3. Core concepts — jobs, scoring, tenancy, substrates
  4. Authentication — credentials

Reference

NeedPage
Every operationAPI Reference
Python clientSandboxClient
Status codes and quotasErrors and limits
PostmanPostman

Machine-readable contracts

ArtifactPath
OpenAPI/openapi/sandbox-platform.yaml
This documentation/docs