Skip to main content
Beam (beam.cloud) is a serverless CPU/GPU cloud whose Sandboxes are disposable containers for agent code that boot in seconds and can fork from snapshots. The following table shows what CostGraph runs on this provider. CI runners: no. AutoPilot compares runners by price, and this provider publishes no sandbox price.

​
Connect

Run the dispatcher with provider set to beam and these variables:
  • BEAM_TOKEN
  • BEAM_WORKSPACE_ID

​
Key and permissions

BEAM_TOKEN is a Beam workspace API token, created in the Beam dashboard under Settings > API keys (platform.beam.cloud/settings/api-keys). BEAM_WORKSPACE_ID is the workspace’s external_id, returned by GET https://app.beam.cloud/api/v1/workspace/current with that token. Beam API tokens are workspace credentials with no per-token scopes or roles. Sources: docs.beam.cloud: authentication, docs.beam.cloud: platform api, docs.beam.cloud: configuration, docs.beam.cloud: overview, docs.beam.cloud: pricing and billing.

​
Limits

  • Sandbox lifetime is keep_warm_seconds, a TTL from creation, not an idle timer: ‘Running processes do not make a finite TTL unlimited.’ Python default is 600 seconds; -1 means run until explicitly terminated; sb.update_ttl() resets it. (source)
  • Python Sandbox defaults to 1 CPU and 128 MiB memory; larger cpu/memory and GPUs are requested per sandbox. No per-sandbox maximum is documented. (source)
  • Billing is usage-based; prepaid workloads stop when the credit balance runs out. (source)