Skip to main content
Velqa
← Back to home
beta

Velqa Sandbox Agent — run code in an ephemeral Docker sandbox (BETA)

Velqa's Sandbox Agent lets you build AI agents that actually run code inside an isolated, ephemeral Docker sandbox, straight from the dashboard. This is a BETA feature, open on every plan (Starter, Dev, Pro).

The model: one ephemeral sandbox per turn

Unlike plain chat, where the model only produces text, a Sandbox Agent gets a real execution environment:

  • On every conversation turn, Velqa spins up a fresh Docker sandbox (Python or Node, depending on the agent).
  • Inside it, the agent can write and read files, run shell commands, execute code, keep Python state across calls, and publish a deliverable (see agent tools). The web tools (web_search, web_fetch) are documented but not enabled yet.
  • The sandbox is isolated and ephemeral: it stays alive for roughly 10 minutes while a turn is in flight, and is released about 2 minutes after the agent stops working — then disappears. Nothing persists from one sandbox to the next except what you keep via checkpoints or publish as an artifact.

Access, billing and limits

  • Open on every plan — Starter, Dev and Pro. There is no allowlist any more; the only gate is fleet capacity, see below.
  • Compute is billed at $0.005 per minute of sandbox, measured on wall clock from creation to release. This is a real charge, not an estimate.
  • Every plan gets a monthly included allowance: roughly 1 h on Starter, 5 h on Dev, 15 h on Pro ($0.30, $1.50 and $4.50 of compute respectively). Past the allowance, minutes are drawn from your Boost balance.
  • A daily spend cap bounds the damage: $1 on Starter, $2 on Dev, $4 on Pro. A turn whose estimated worst-case cost would push you past the day's cap is refused before it starts.
  • One sandbox at a time per account, on every plan.
  • The platform fleet is bounded. When every slot is taken, creation returns a 429: this is neither your rate limit nor your balance — retry in a few minutes.

When to use it instead of plain chat

Reach for the Sandbox Agent when you need the AI to do something, not just describe it:

  • run and verify code (Python/Node), run tests, manipulate files;
  • process a file you upload into the workspace;
  • chain several tool-driven steps (search, compute, produce a downloadable deliverable).

For simple Q&A or text generation, stay on chat: it's faster and cheaper.

Next steps