It seems like OP needs to provide a solution to hiding the credentials from the model in order to suggest CLI-mode only, and also a solution to the problem of agents without shell access.
Instead of forwarding ANTHROPIC_API_KEY / OPENAI_API_KEY into the guest, we simply run a reverse-proxy within the VM that replace a constant-time token with the real API key on the host.
Works well enough in practice
I think something like infiscial ai proxy could be useful here. Never store the creds on device.
(leaving out cases where your genius GPT-12 Galaxy Ultra agent hacks the sandboxing from inside)
This is the biggest problems with most “sandboxes”. Some people aren’t even running a sandbox. But even the best have a big problem: APIs where GET verbs provide write features.
This is the value of MCP: minimize the surface to known APIs and identify read-only from mutating so I can trust, approve or block. The MCP server, in this case, does NOT run in an environment that the read/write or shell can see.