Skip to content

Run commands & processes

mvm has two command styles:

one-shot sandboxes that boot, run a command, and exit;
commands inside an already-running named microVM.

Use one-shot mode for isolated code execution. Use named-VM commands when you are working with persistent state, services, logs, or cold-mode recovery.

One-shot run

mvmctl run -- uname -a
mvmctl run --profile restrictive -- python -c 'print("hello")'
mvmctl run --timeout 30 --receipt /tmp/run-receipt.json -- python task.py

mvmctl run produces a transient sandbox. It can write a signed receipt with invocation hashes, output hashes, and exit status. Raw argv, env values, stdout, stderr, and host paths are not stored in the receipt.

Use dry-run mode to inspect policy effects without booting:

mvmctl run --dry-run --json -- python task.py

Named VM commands

mvmctl up ./my-app --name agent-sandbox
mvmctl exec agent-sandbox -- python /work/task.py
mvmctl logs agent-sandbox -f

Use this path when the VM has state, files, services, or snapshots that should survive across commands.

Profiles

Profile	Use it when	Notes
`restrictive`	Running generated or untrusted code.	No env injection and no host directory shares.
`standard`	Normal local runs.	Explicit env is allowed; host shares must be read-only.
`dev`	Iterating on a local project.	Writable host shares are allowed.
`permissive`	Last-resort debugging.	Requires explicit acknowledgement.

Security notes

Prefer argv arrays and explicit command arguments.
Avoid passing secrets through command-line args.
Use receipts for automation and audit correlation.
Keep writable host shares out of untrusted runs.
Treat stdout and stderr as sensitive because guest code controls them.