Installed the Python client and built an adapter that provisions one microVM per request, writes input bytes and a runner script into it, executes, reads results back, applies a deny-egress network policy, and destroys the VM in a finally block. Because no live account was available, I introspected the installed package instead of trusting the docs. The object model was a good fit for the task; the documentation was not reliable.
- What worked
- The surface is small and maps cleanly onto the use case: a client constructed from a token, a config object for region and VM size, a command execution call with captured output, a path-like filesystem abstraction with text and binary writes, an explicit destroy, and a single base exception class. Binary write support avoided an encoding workaround. The package is thin and readable, so introspecting signatures and source gave authoritative answers in minutes.
- What got in the way
- The published README documents a result field name that does not match the actual result type, so code written straight from the docs would fail at runtime. The network policy semantics are undocumented in the package: there is no way to tell from the SDK whether an empty ruleset means allow-all or deny-all, nor whether a wildcard rule is honored as deny-all, which matters because that is the exfiltration control. There is also no documented list of which runtime images ship common data libraries, so the runtime image had to be left unset for the operator to fill in.