Provider Landscape
Read when:
- comparing Crabbox with adjacent sandbox, dev-environment, or microVM tools;
- deciding whether a competitor should become a first-class provider;
- planning provider capability work without live provider credentials.
Use this architecture map to compare execution contracts. It is not a benchmark; check the linked provider documentation for current behavior.
#Core Position
Crabbox should own the provider-neutral contract: selecting a target, syncing a checkout, running commands, returning proof, exposing reachability, and cleaning up capacity. It should not copy every adjacent product's control plane.
That means:
- support stable execution substrates as providers;
- expose shared workflow capabilities through
crabbox providersandcrabbox providers recommend; - keep runtime-shape labels provider-neutral (
managed-sandbox,delegated-command,ssh-host,worker-module) instead of copying one vendor's control-plane terms; - keep reachability labels provider-neutral (
provider-url,ssh-tunnel,tailnet-peer,tailnet-egress) so preview URL, tailnet, and SSH tunnel workflows can be compared without provider-specific docs scraping; - use
sshorexternalfor tools that already produce a normal host contract; - defer first-class support for runtimes whose useful behavior is still too product-specific to test offline.
#Landscape
| Segment | External systems | Crabbox fit today | Improve next |
|---|---|---|---|
| CI proof runners | Blacksmith Testbox, Semaphore, GitHub Actions-style runners | Strong fit. Crabbox already separates proof-runner semantics from generic VM semantics through ci-proof-runner providers and providers recommend ci-proof. | Keep run proof, artifacts, downloads, and status records normalized across providers. |
| Hosted agent sandboxes | E2B, Vercel Sandbox, Modal Sandboxes, Cloudflare Sandbox SDK, OpenSandbox, smolvm, Upstash Box | Good fit when the provider owns process execution and files. These map to delegated-run better than SSH leases. | Improve artifact/download parity, preview URL reporting, timeout/error taxonomy, and optional MCP attachment routing. |
| Remote developer environments | Coder, Daytona, Namespace Devbox, CodeSandbox, Morph, OpenComputer, Codespaces-like tools | Good fit when Crabbox can either SSH into the workspace or delegate a command with archive sync. providers recommend remote-dev is the routing surface. | Add clearer live smoke docs per provider, surface pause/resume support, and keep local-editor, remote-compute flows distinct from CI proof. |
| Forkable/versioned workspaces | Mitos, Firecracker snapshot systems, local-container, Parallels | Crabbox has provider-neutral checkpoint/fork/restore capabilities and checkpoint fan-out via checkpoint fork --count. The catalog advertises workspace fork on Daytona, Hetzner, Incus, local-container, Machine0, and Parallels; these are not uniform live-memory forks. | Harden versioned-workspace behavior before adding any runtime-specific fork API. Do not add Mitos-only flags. |
| Worker and module runtimes | Cloudflare Dynamic Workers, Cloudflare Sandbox SDK, Vercel/edge-adjacent runtimes | Narrow fit. cloudflare-dynamic-workers is a module-run provider; generic container sandboxes need a separate lifecycle and file/process contract. | Keep worker-runtime separate from Linux sandbox execution unless the provider can expose files, process status, logs, preview URLs, and cleanup. |
| Self-hosted virtualization | Proxmox, XCP-ng, Incus, KubeVirt, local VMs | Strong fit when Crabbox gets a normal SSH lease and lifecycle hooks. | Keep provider-specific reconciliation behind adapters; no provider-specific branching in core. |
| GPU and ML execution | RunPod, NVIDIA Brev, W&B, Modal, cloud GPU VMs | Mixed fit. SSH leases are best for debugging; delegated providers are best for provider-owned jobs. | Normalize result evidence and cost/usage metadata before chasing more GPU-specific launch paths. |
#Mitos Decision
Do not add first-class Mitos support yet.
Mitos targets forkable Firecracker microVMs, MCP, and copy-on-write fan-out. A Crabbox adapter would need to keep its Kubernetes/KVM requirements behind the provider boundary and prove the lifecycle on a real cluster.
Develop the shared contracts first: workspace-checkpoint, workspace-fork, workspace-restore, provider-snapshot, mcp-attachments, and portable run evidence (run-proof, run-artifacts, run-downloads, url-bridge). The runtime, lifecycle, and reachability filters already describe execution and access without vendor-specific CLI modes. fanout-testing (alias best-of-n) selects fork-capable providers; checkpoint fork --count fans out archive or native checkpoints without promising live-memory cloning.
Revisit Mitos when a concrete workload needs live microVM forking and the adapter has offline coverage plus an opt-in live smoke.
#Support Matrix
| System | Support stance | Why |
|---|---|---|
| Mitos | Observe, do not support directly yet. | Its main differentiator is live microVM forking. Crabbox needs generic fork/checkpoint semantics hardened before a Mitos adapter would be clean. |
| E2B | Supported as e2b. | Delegated sandbox execution maps cleanly to provider-owned sessions and URL bridge behavior. |
| Vercel Sandbox | Supported as vercel-sandbox. | Ephemeral Linux sandbox execution fits delegated-run; keep adding evidence and artifact parity before more surface area. |
| Modal | Supported as modal. | Provider-owned container execution fits delegated runs, especially Python and ML-shaped workloads. |
| Cloudflare Sandbox SDK | Supported through cloudflare-sandbox. | Delegated Linux commands and archive sync use an operator-deployed Sandbox bridge; separate from Cloudflare Containers and Dynamic Workers. |
| Cloudflare Dynamic Workers | Supported as cloudflare-dynamic-workers. | Worker/module execution is a separate target from Linux sandbox execution. |
| Daytona | Supported as daytona. | Direct ordinary runs use SDK/toolbox APIs; script runs and brokered execution use SSH. |
| Namespace Devbox | Supported as namespace-devbox. | SSH lease plus sync/cleanup behavior fits remote-dev routing. |
| CodeSandbox | Supported as codesandbox. | Delegated workspace execution and pause/resume fit remote-dev routing. |
| Morph | Supported as morph. | Managed SSH lease fits local-editor, remote-compute workflows. |
| OpenComputer | Supported as opencomputer. | Delegated Linux execution fits remote-dev and sandbox routing, but evidence parity should improve. |
| DevPod | Do not support directly. | It is already a provider-agnostic dev environment layer. Use the resulting SSH/container target through ssh, local-container, or external. |
| Coder | Supported as coder. | The built-in contract is intentionally narrow: local Coder CLI auth, Linux SSH proxy execution, stop-by-default release, delete only by opt-in, and cleanup only for locally claimed workspaces. Use ssh or external for existing workspaces that Crabbox should not manage. |
| Kubernetes Agent Sandbox | Supported as agent-sandbox. | SandboxClaim-style delegated execution belongs behind the existing Kubernetes-native adapter. |
| OpenSandbox | Supported as opensandbox. | Delegated execution through compatible OpenSandbox control and exec endpoints, with archive sync and owned sandbox cleanup. |
| Microsandbox, Moru-like runtimes | Candidate only. | Add only when one has a stable lifecycle and evidence contract that is meaningfully different from existing delegated sandboxes. |
#Roadmap
- Keep recommendations centered on workflows such as
ci-proof,run-evidence,fast-feedback,remote-dev, andteam-cloud. Use the command reference for the full list, aliases, filters, and scoring. - Improve delegated-provider evidence parity: proof, artifacts, downloads, preview URLs, retained logs, reusable sessions, and error status. Access and evidence metadata should describe what users can inspect after a run.
- Document an opt-in live smoke for each provider, including prerequisites and verified cleanup. Offline recommendations do not establish working credentials, capacity, or a security boundary.
- Strengthen checkpoint, fork, restore, and snapshot semantics before adding runtime-specific forking.
workspace-reuseandfanout-testingexpose these workflows without assuming live-memory cloning. - Keep
worker-runtimemodule execution distinct from Linux shells and filesystem sandboxes in both targets and feature declarations.
#Add-Provider Bar
Follow Provider authoring for backend selection, capability declarations, credential handling, tests, and documentation. A new adapter must prove create, inspect, run or attach, release, and cleanup behavior under the live smoke contract. Use ssh or external for a manual integration until that contract is ready.