Ollama Cloud now works inside Fly.io Sprites, full Linux virtual machines built for agents. Agents running in a Sprite can call cloud models such as kimi-k3, glm-5.3 and deepseek-v4.1-flash without ever holding your Ollama API key.

For anyone sandboxing coding agents, this means model access without putting credentials inside the sandbox.

Key Features at a Glance

• Full Linux sandboxes: Each Sprite is a Linux VM with its own disk that persists between sessions, so agents can run commands securely in a contained environment.
• Agent support: Claude Code, Codex and Gemini CLI are supported out of the box.
• Key protection: Fly.io encrypts your Ollama key and never returns it through the API. A gateway adds the key to requests, so an agent in a Sprite can't leak it.
• Default-deny access: New connectors block every Sprite until you grant access by label, by name prefix, or to all Sprites.
• Inside-only calls: Ollama calls work only from inside a Sprite.
• Agent skill: Claude Code, Cursor, Codex and Gemini CLI get a skill that teaches them about Ollama, so permitted agents can use its models without extra setup.

Availability

The Ollama Connector is set up in the Fly.io dashboard under Sprites and then Connectors, and requires an Ollama API key and at least one Sprite; the article was published October 13, 2026.

Fly.io Sprites are persistent, hardware-isolated Linux computers implemented as Firecracker microVMs, each with its own kernel and durable filesystem. They provide up to 8 CPUs and 16 GB of RAM per run, a 100 GB durable root filesystem, automatic idle sleeping, HTTPS access through an individual URL, and checkpoint-and-restore for the full environment. Sprites can be created through Fly.io’s CLI, REST API, or SDKs, and new instances start in seconds.