Product surface

grain’s local microVM product surface as of the v0.2 line.

Complete for local sandboxes

Area	Status
Daemon + unix/TCP API + optional auth	Done
Ephemeral / persistent lifecycle	Done
Guest agent (exec, shell PTY, cp, fs, stats, secrets)	Done
Create wait modes (auto/ssh/agent/userdata)	Done
Golden `grain-ubuntu` pull/bake/import	Done
Multi-distro: ubuntu-cloud, alpine-cloud	Done
Port forwards (create + live)	Done
Mounts 9p; virtiofs on Linux	Done
Pause/resume; suspend/restore	Done
Resource caps, profiles, presets	Done
Egress proxy (default-deny + secret inject)	Done
Go / TypeScript / Python SDKs + OpenAPI	Done
MCP server (`grain up --mcp` / `grain mcp`) for agent hosts	Done
Install script, recipes, create bench	Done
Guest arch selection (`--arch`, incl. x86_64 on Apple Silicon via QEMU)	Done
Virtio GPU (`--gpu` / `gpu: virtio`)	Done
Shared overlay network between VMs (`network: overlay`)	Done
Firecracker backend	Experimental (Linux)

Platforms: macOS and Linux hosts with hardware virtualization. Native Windows is not a host (use the remote API/SDKs against a supported host). WSL is Linux from grain’s point of view; virt must be available to that environment.

Quick verification

bash

just test && just smoke-api
just build && just agent-linux
./bin/grain doctor
# live (optional):
# ./bin/grain up && ./bin/grain image pull grain-ubuntu && ./scripts/bench-create.sh

Create latency

Measure on your hardware (numbers vary by image, wait mode, and host):

bash

grain up
grain image pull grain-ubuntu
./scripts/bench-create.sh -n 5 --wait agent

Use the reported p50/p95 when talking about boot speed — do not invent sub-300ms claims without a local run.