Sandbox API

Sandbox API

Run agents, benchmarks and untrusted code in isolated sandboxes over one HTTP API.

Sandbox runs code and AI agents in isolated environments and gives you back the results. You send a request; the platform provisions a sandbox, runs the work, scores it if there is a verifier, and keeps the 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 own host

Choose an endpoint

Pick by what you are trying to do. All four are available to every credential.

I want to…EndpointPython client
Run a benchmark task and score itPOST /sandbox/harbor/v2/jobs/execute-tasksrun_datapoint(...)
Run a coding agent against a repoPOST /sandbox/agent/v1/runsrun_agent(...)
Run a snippet of code once, no agentPOST /sandbox/codeedit/v1/executionsrun_code_execution(...)
See which agents and languages existGET /sandbox/v1/catalogcatalog(kind=...)

If you need a score, a verifier, or a task archive, you want the first one. If you just need output from code, you want the third.

Start here

  1. Get started — your first request, end to end
  2. Architecture — components + how requests flow
  3. Core concepts — jobs, statuses, artifacts, tenancy
  4. Authentication — credentials and what they can reach

Reference

Every endpoint, with parameters, schemas and examples:

Machine-readable contracts

Served from the same hostname as the API, so these paths work in every environment.

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

Limits worth knowing up front

  • Concurrent runs are capped per project. Over the cap, synchronous requests get 429 with Retry-After; queued jobs simply wait.
  • Task archives are capped at 256 MiB compressed and 1 GiB expanded.
  • Job records and artifacts expire, so download what you need rather than treating the platform as long-term storage.

Details and every error code: Errors and limits.