CrowNest
Guides

CrowNest vs direct Cloudflare Sandbox

Use CrowNest when agents need durable evidence, scoped product controls, and replayable run records on top of the runtime primitive.

Cloudflare Sandbox is the runtime primitive. CrowNest is the product layer for agents and applications that need that runtime to leave behind durable, queryable proof.

Use CrowNest when the output of a run must be inspectable after the sandbox is gone: an exit code, stdout and stderr events, explicit artifact IDs, archive checksums, source metadata, timing, and an Evidence Bundle that other agents or CI jobs can read later.

What CrowNest adds

NeedDirect runtime callCrowNest surface
Durable run proofCaller stores logs and exit codesWorkspace Run evidence persists terminal proof.
Artifact retentionCaller copies files out before cleanupArtifacts are explicit retained resources.
Auth and scopingRuntime credential handlingOrg, project, and API-key scopes gate every public operation.
Idempotent remote workflowsCaller defines retry recordsWorkspace Runs and uploads accept idempotency keys.
Agent replayCaller builds event storageEvents can be streamed live or replayed by sequence.
Dashboard inspectionCaller builds the operator surfaceRuns, previews, artifacts, and usage are visible in the dashboard.
Usage and quotasCaller meters product actionsCrowNest records billable usage and enforces plan/project limits.

Evidence-first workflow

For repository-sized work, create a Workspace Run from an archive, stream events for progress, then read the Evidence Bundle as the durable result.

Terminal
crownest workspace-runs run-archive repo.tgz --template python-node -- pnpm test
crownest workspace-runs evidence wsr_abc123 --output evidence.json

The event stream is useful while the command is running. The Evidence Bundle is the handoff object after the run is terminal: it records status, exit code, archive checksum, artifact IDs, artifact collection errors, timing, metadata, and failure details.

When direct runtime calls are enough

Call the runtime directly when you own the whole execution loop and do not need a product record afterward. That can be right for short-lived internal tools, runtime experiments, or one-off scripts where stdout in the caller process is the only output.

Use CrowNest when another person, job, or agent needs to answer "what happened?" after the environment has been cleaned up.

On this page