Ollama Announces Integration with Fly.io Sprites

At a Glance
| Item | Value |
|---|---|
| Publisher | Ollama |
| Published | 2026-10-13 |
| License | MIT |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code when the information was collected. Dates are JST.
Overview
On October 13, 2026, Ollama announced an integration with Sprites, a sandbox environment provided by Fly.io. This partnership makes it possible to utilize the latest open-weight models provided by Ollama Cloud while running AI agents safely inside fully functional Linux virtual machines. Equipped with a secure connection mechanism that prevents API key leakage, it allows developers to have agents execute advanced tasks while maintaining confidentiality.
Announcement Details
At the core of this announcement is the new “Ollama Connector," which bridges Fly.io Sprites and Ollama Cloud. Sprites is a sandbox environment consisting of complete Linux virtual machines (VMs) with dedicated disks that persist data between sessions. Using this environment, agent tools such as Claude Code, Codex, Gemini CLI, and Cursor can be operated in a secure state isolated from the host system.
Secure Connection Mechanism
The Ollama Connector adopts a mechanism where API keys are not placed directly in the environment where the agent runs. Users first add an Ollama API key to their organization on the Fly.io dashboard. This key is encrypted on the Fly.io side and is never read externally via the API.
When calling Ollama models from inside a Sprite, it goes through a gateway provided by Fly.io. The gateway identifies the Sprite that sent the request, automatically attaches the API key, and forwards it to Ollama Cloud. Since only the gateway URL and connection ID (connection_id) are passed to the Sprite side, this structurally eliminates the risk of agent code leaking API keys externally, either accidentally or maliciously.
Access Control and Flexible Configuration
To further strengthen security, the new connector is designed by default to block access from all Sprites. Administrators can finely control which environments are granted access rights to Ollama in units such as labels, name prefixes, or all Sprites. In addition, this communication is valid only from inside the Sprite, designed to prevent unauthorized use from the outside. After configuration, the “Test reachability" feature on the dashboard can be used to verify whether a specific Sprite can correctly connect to Ollama.
Supported Models and Interfaces
Through this integration, the latest models on Ollama Cloud such as kimi-k3, glm-5.3, and deepseek-v4.1-flash become available. Usage methods are extremely flexible; in addition to access via standard curl commands, they can also be called from Python code compatible with the OpenAI SDK. When using Python, it works simply by setting base_url to the gateway URL and specifying any value (such as “unused") for api_key. Ollama’s own /api/* routes are also supported via the gateway. Since Fly.io Sprites have built-in features to inform Claude Code, Cursor, Codex, and Gemini CLI of Ollama’s presence, models can be used from these agents without additional configuration if proper access permissions are in place.
Practical Use Case: Fuzzing
The announcement cites “fuzzing" (vulnerability assessment) as a specific use case combining multiple models. For example, a relatively small model with 30 billion parameters, such as nemotron-3-nano:30b, is made to generate malicious inputs, which are then executed against a program inside the Sprite. A workflow where behaviors such as crashes generated as a result are analyzed by a larger and more advanced model like kimi-k3 is presented. Since switching models can be easily done just by changing the model specification field within the request, using models according to task difficulty is recommended.
Background
The documentation explicitly states that this integration was developed jointly by both the Ollama and Fly.io teams. At the end of the blog, expressions of gratitude are offered with “Thank you to the Fly.io team for building this with us," indicating that this was not merely a one-sided partnership announcement, but a mechanism built through cooperation between the two companies.
The documentation also touches upon the point that coding agents such as Claude Code, Codex, Gemini CLI, and Cursor are designed on the premise of running within sandboxes when executing commands. It can be inferred that the fact these agent tools have acquired command execution capabilities made the isolation of execution environments and the protection of API keys an issue, although the documentation does not explain the background circumstances in further detail.
Impact on Local LLM Users
First, it should be noted that this integration targets “Ollama Cloud" and not the local Ollama server running on the user’s own machine. kimi-k3, glm-5.3, and deepseek-v4.1-flash mentioned in the documentation are all introduced as models provided via Ollama Cloud, and agents running inside Sprites call these models across the network. Readers who assume the use case of running inference on their local GPUs need to understand that they cannot bring the mechanism introduced here directly into their own local environments.
On the other hand, the models provided on Ollama Cloud themselves are open-weight models, which is consistent with the fact that the ollama/ollama project supports models like Kimi, GLM, and DeepSeek. In other words, while the option itself to run the same model family locally exists separately, there is no clear mention in the documentation as to whether this Sprites integration supports that local execution path.
On the security front, it is explicitly stated that code within the Sprite does not hold API keys itself, with a mechanism where the gateway attaches keys on its behalf. For readers who want to let agents process advanced tasks while granting them access to the external network, but want to avoid the risk of key leakage, this is a design pattern worth referencing. However, it cannot be overlooked that using this mechanism presupposes registering for Fly.io’s Sprites service and having a usage contract for Ollama Cloud (obtaining an API key). The documentation also shows a reference link to “Ollama pricing," suggesting a configuration where billing may occur for Cloud-side usage, though specific pricing is not stated in this documentation and is noted as “not mentioned."
Additionally, licenses and benchmark scores are not described in this documentation, and mentions of these points are avoided. For the parameter count of nemotron-3-nano:30b (30 billion) cited in the fuzzing example, only the numeric value explicitly stated in the documentation is adopted.
Related Articles
- Cloudflare’s Clef Multimodal Decision Model Now on Ollama
- Ollama v0.35.0 Released: New API for Decision Models
What to Read Next
- Engines mentioned in this article → Ollama

