Provider Selection
Read when:
- choosing where a workload should run;
- comparing a Crabbox provider with an external sandbox or dev-environment tool;
- deciding whether a new provider belongs in Crabbox.
For a deeper competitor and roadmap view, read Provider landscape.
Crabbox should support provider families, not every interesting runtime by name. Prefer the provider that fits the workflow and its evidence requirements. Add a new first-class provider only when the runtime exposes a stable lifecycle, execution, authentication, cleanup, and proof contract that Crabbox can test without special operator knowledge.
#Fast path
Start with the command-backed recommendation list:
crabbox providers recommend
crabbox providers recommend artifact-download
crabbox providers recommend ci-proof
crabbox providers recommend cost-control
crabbox providers recommend linux-vm --limit 8
crabbox providers recommend agent-sandbox --json
crabbox providers recommend live-smoke
crabbox providers recommend mcp-sandbox
crabbox providers recommend network-isolation
crabbox providers recommend pause-resume
crabbox providers recommend preview-url
crabbox providers recommend remote-dev
crabbox providers recommend run-session
crabbox providers recommend workspace-reuse
crabbox providers recommend forkable-workspace --workspace fork
The recommendation command uses the built-in provider spec and checked-in selection metadata only. It does not contact live providers, inspect quota, or prove credentials. Treat it as routing advice, then run:
crabbox doctor --provider <name>
before spending real capacity. For the live proof expected from provider PRs, see Provider live smoke.
Lifecycle commands require actionable provider intent: pass --provider <name>, set CRABBOX_PROVIDER, or configure a provider in user, repository, or broker config. The compiled provider value is retained only for compatible default derivation and is never enough to initialize a backend. Existing lease commands may instead route from an exact local claim, an unambiguous claimed slug, or recorded run context. Exact claim and recorded routes are authoritative: only an explicit routing flag may replace them. Archive checkpoint-fork and shard dry-runs still require provider intent because the archive does not record a destination provider. With no actionable selection, commands stop with shared guidance before provider credential validation. Run crabbox providers recommend when choosing a provider.
#Selection rules
Use these rules before adding a new adapter:
- If the provider can expose a stable SSH target, prefer an SSH-lease backend. That keeps sync, Actions hydration, VNC, code-server, cache, result parsing, and downloads in core.
- If the provider owns filesystem sync and command execution, use a delegated run backend and declare only the features the provider actually supports.
- If the provider only starts, stops, or inspects an existing service, use a service-control provider. Do not route arbitrary
crabbox runthere. - If the integration depends on local scripts or a private fleet contract, start with
externalorsshbefore adding built-in provider code. - If the provider cannot be validated without live credentials, add offline unit tests and docs first; keep live proof as an opt-in smoke.
#Workflow map
| Workflow | Prefer | Why |
|---|---|---|
| CI reproduction with durable proof | blacksmith-testbox, semaphore | They map to CI/proof-runner semantics instead of generic devbox semantics. |
| Run artifacts and downloads | blacksmith-testbox, islo | They advertise artifact or download evidence and appear in providers recommend artifact-download. |
| Run evidence and previews | blacksmith-testbox, islo, e2b | They advertise normalized proof, artifact, download, or preview-url capabilities in crabbox providers and providers recommend run-evidence. |
| Inspectable run sessions | local-container, blacksmith-testbox, islo, e2b, cloudflare-dynamic-workers | They advertise reusable session evidence and appear in providers recommend run-session. |
| Pausable or resumable runtimes | codesandbox, islo | They advertise pause/resume capability and appear in providers recommend pause-resume. |
| Provider preview URLs | islo, e2b, railway | They advertise provider-native preview URL evidence and appear in providers recommend preview-url. |
| Provider live smoke | blacksmith-testbox, apple-container, local-container, cloudflare, aws, sealos-devbox | They advertise enough sync, cleanup, lifecycle, or evidence capability for providers recommend live-smoke to rank them as good opt-in smoke candidates, with local runtimes kept visible when cloud credentials are unavailable. |
| Cost- or quota-sensitive validation | local-container, apple-container, multipass, aws, azure | They advertise local execution, cache reuse, cleanup, or coordinator-governed cloud controls and appear in providers recommend cost-control. |
| Fast feedback with reusable caches | local-container, apple-container, multipass, blacksmith-testbox | They advertise cache-volume, sync, cleanup, or reusable proof/session capabilities in crabbox providers and providers recommend fast-feedback. |
| Disposable isolated execution | agent-sandbox, anthropic-sandbox-runtime, e2b, smolvm, vercel-sandbox | They are delegated or local sandbox providers in crabbox providers and providers recommend isolated-execution. |
| Network-contained untrusted execution | cloudflare, agent-sandbox, codesandbox, opensandbox, vercel-sandbox | They keep execution inside delegated or local sandbox boundaries and appear in providers recommend network-isolation. |
| MCP-attached sandbox runs | docker-sandbox | It advertises mcp-attachments in crabbox providers and providers recommend mcp-sandbox. |
| Remote developer environments | namespace-devbox, daytona, morph, sealos-devbox, codesandbox, opencomputer | They are managed dev-environment substrates with checkout sync, SSH or provider-owned workspace access, and providers recommend remote-dev routing. |
| Shared app reachability | hetzner, azure, gcp, islo, e2b | They advertise tailnet, URL bridge, or SSH tunnel planes in crabbox providers and providers recommend reachability. |
| Shared team cloud leases | aws, azure, gcp, hetzner | They advertise brokerable cloud, cleanup, SSH, and sync capabilities in crabbox providers and providers recommend team-cloud. |
| Generic Linux command execution | aws, azure, gcp, hetzner, digitalocean, linode, ssh | SSH leases keep the normal Crabbox sync/run/debug path. |
| Existing owned machine | ssh | No provider lifecycle is needed; Crabbox only syncs and runs. |
| Local disposable Linux | local-container, apple-container, apple-vm, multipass | Fast local iteration without cloud credentials. |
| Native desktop/browser/code-server | aws, azure, hetzner, parallels, local-container, ssh | These advertise the interactive lease features. |
| GPU-oriented run | runpod, nvidia-brev, cloud VM providers with GPU types, modal, wandb | Pick SSH leases for normal debugging, delegated runs for provider-owned ML execution. |
| Worker/module execution | cloudflare-dynamic-workers | It advertises the worker-runtime target and module-run feature. |
| Versioned workspace reuse | parallels, local-container | They advertise normalized checkpoint/fork/restore/snapshot-reference capabilities in crabbox providers and providers recommend versioned-workspace; workspace-reuse and forkable-workspace are aliases for the same workflow. |
| Self-hosted virtualization | proxmox, xcp-ng, incus, kubevirt, external, ssh | Keeps private infrastructure behind explicit provider boundaries. |
#Related external systems
The Provider landscape records support status and comparisons with adjacent systems. Provider pages document the implemented contracts. Prefer ssh or external for an existing host or a private lifecycle integration that does not need a built-in adapter.
#Mitos decision
Mitos remains unsupported. See the landscape decision for the rationale and provider-neutral checkpoint, fork, and evidence capabilities to develop first. There is no provider: mitos or Mitos-specific CLI mode.
#Support threshold
A built-in provider needs an accurate execution kind, target and capability metadata, off-argv authentication, explicit ownership and cleanup, offline regression tests, and a documented live smoke. Follow Provider authoring and the live smoke contract.