Ollama Cloud Connector Launches for Fly.io Sprites to Secure Agent Execution
Fly.io's persistent Linux sandbox VMs now route calls to Ollama Cloud models like kimi-k3 and glm-5.3 through an encrypted gateway that injects API keys without exposing them.…
- Sprites function as full Linux virtual machines with persistent disks, allowing agents to run commands safely while Ollama requests are proxied through a Fly.io gateway that authenticates the caller and adds the encrypted key.
- Connectors start with all Sprites blocked; owners grant access selectively by label, name prefix or individually, and calls are restricted to originate only from approved sandboxes.
- OpenAI-compatible SDKs, native Ollama API routes and pre-configured skills for coding agents like Claude Code and Gemini CLI enable seamless model switching with no additional setup inside authorized Sprites.
Gateway-Based Key Protection
The Ollama Connector prevents credential exposure by keeping API keys out of Sprite environments. Users add their key once in the Fly.io organization dashboard where it is stored encrypted.
When code inside a Sprite calls Ollama, the request travels to a gateway that verifies the originating Sprite via connection ID and injects the key server-side. The Sprite only knows a gateway URL and never receives the actual credential.
Configuration and Usage
Setup begins with an Ollama API key and at least one Sprite. In the dashboard under Sprites then Connectors, owners create an Ollama connector, configure access rules and save. A test reachability button verifies connectivity from each Sprite.
Inside an authorized Sprite, standard curl commands or any OpenAI SDK can target the gateway endpoint using a placeholder api_key value. Native Ollama /api routes are also supported through the same proxy. The connector page provides a ready-to-run example command with the correct connection ID.
Agent Tooling and Practical Examples
Sprites ship with built-in skills that register Ollama models with tools such as Claude Code, Cursor, Codex and Gemini CLI. Authorized agents can invoke models directly by updating the model parameter in requests.
One documented workflow uses multiple models for fuzzing: a compact model generates malformed inputs, the Sprite executes them against a target binary, and a larger model examines any crashes. Changing the model name switches between available options like nemotron-3-nano and kimi-k3.
Availability and Next Steps
The feature is available immediately to users with an Ollama API key and Fly.io Sprites. Documentation includes a quickstart guide on the Fly.io site covering connector creation, access controls and example code.
This integration offers developers a practical way to combine sandboxed Linux execution with hosted models while maintaining credential isolation. Builders should review the official quickstart to assess fit for their agent workloads.
