Ollama Cloud now works from Fly.io’s persistent Linux Sprites. The connector stores the Ollama key in the Fly.io organization, and a gateway adds it to authorized requests without exposing it directly to agent code.
Each Sprite has its own virtual machine and persistent disk. New connectors deny all Sprites by default; administrators select access by name, prefix or label. OpenAI-compatible and native Ollama routes go through the gateway. Accounts and resources from both services remain necessary.
Keeping the key outside the sandbox reduces one secret-leak path but does not prevent abuse of authorized access or exfiltration through other channels. These are cloud models, not a new free local-inference release.
Credentials can be separated from untrusted code execution without removing the need for action governance.