Ollama Models Now Run Inside Fly.io Sprites
Ollama published a blog post on October 13, 2026 announcing that Ollama Cloud models can be used from inside Fly.io Sprites through a new Ollama Connector. Everything below comes from Ollama's own page.
Transcript
Ollama says an AI agent can use its hosted models without holding credentials. Fly stores them instead.
The post says Ollama Cloud models now run inside Fly's Sprites, sandboxed Linux machines with persistent disks.
Ollama states that Fly encrypts the key you provide and never returns it.
Ollama writes that a gateway identifies the caller and attaches the secret. The sandbox gets only a connection identifier.
The company claims a sandboxed agent cannot leak what it never receives. New connectors block every sandbox by default.
This is a first party announcement from interested companies. No pricing or performance figures appear.
Inboxsmith tracks the claims behind company announcements. Please like and subscribe for more news.
Sources
Every claim in this video comes from the top ranking coverage of this topic. The claims and where each one came from:
- Line by line backing, all of it from the single Ollama post. Hook sentence one, "Ollama says an AI agent can use its hosted models without holding credentials": supported by "With the new Ollama Connector, your Sprites can use Ollama's cloud models" together with "An agent in a Sprite can't leak your key" and "The Sprite itself only knows a gateway URL and a connection ID"; the phrase "hosted models" was chosen over "cloud models" to avoid repeating a distinctive phrase that scene one also needs. Hook sentence two, "Fly stores them instead": supported by "You add your Ollama API key to your Fly.io organization, and Fly.io encrypts it and never returns it through the API." The hook is deliberately two sentences because the hook card renders a two beat array and a single sentence leaves the second beat undefined. Scene one, "The post says Ollama Cloud models now run inside Fly's Sprites, sandboxed Linux machines with persistent disks": supported by "Ollama Cloud now works inside full Linux computers designed for agents using Fly.io Sprites", by "A Sprite is a sandbox from Fly.io", and by "Each one is a full Linux virtual machine with its own disk that stays around between sessions"; the article's own wording "stays around" is idiomatic and was replaced with the plain word persistent. Scene two, "Ollama states that Fly encrypts the key you provide and never returns it": supported by "Fly.io encrypts it and never returns it through the API", with the acronym at the end of that sentence dropped from narration since spoken acronyms are avoided where a plain word works. Scene three sentence one, "a gateway identifies the caller and attaches the secret": supported by "When code inside a Sprite calls Ollama, it goes through a Fly.io gateway" and "The gateway checks which Sprite is calling and adds your key to the request." Scene three sentence two, "The sandbox gets only a connection identifier": supported by "The Sprite itself only knows a gateway URL and a connection ID", with the gateway URL half of that claim moved to on screen text for time and the abbreviation ID expanded to the plain word identifier. Scene four sentence one, "The company claims a sandboxed agent cannot leak what it never receives": supported by the stated consequence "An agent in a Sprite can't leak your key", kept explicitly as a claim by the company rather than a verified property because no third party audit exists. Scene four sentence two, "New connectors block every sandbox by default": verbatim support from "New connectors block every Sprite by default." Scene five sentence one, "This is a first party announcement from interested companies": the page is Ollama's own blog describing an integration it built with the Fly.io team. Scene five sentence two, "No pricing or performance figures appear": verified absent on the page, which publishes no pricing for the connector, no latency or throughput figures, no benchmark result, no beta, preview or general availability status, no number of users or Sprites, no region availability, no uptime or rate limit figures, no third party security audit or penetration test, and no outcome data for the fuzzing example including no count of bugs found. Deliberate wording decisions on the record. The company Fly.io is spoken in narration as plain "Fly" because a dotted domain brand has never appeared in our spoken narration and is a known synthesis hazard; the full "Fly.io" is used in the on screen kickers, the on screen evidence lines, the title candidates and these provenance notes. The model names kimi-k3, glm-5.3, deepseek-v4.1-flash and nemotron-3-nano:30b appear only in on screen text and never in any narration field, because brand names glued to digits and hyphenated compounds reliably garble in synthesis; narration refers to them generically as hosted models, Ollama Cloud models, a smaller model and a larger model, with the smaller and larger pair used only in the on screen fuzzing line. The word that describes agent driven software and begins with the letters ag is avoided entirely because the voice reliably garbles it; narration says AI agent and agent instead. Deliberate exclusions. Every enumeration in the source was kept out of narration and placed on screen only, since multi item lists measurably raise the spoken word rate: the three named cloud models, the four named coding agents Claude Code, Cursor, Codex and Gemini CLI, the three access control options by label, by name prefix or all of them, and the four step dashboard setup sequence. Also excluded from narration: the curl and Python code examples, the detail that any OpenAI compatible client works and that the native Ollama routes pass through the gateway, the skill that tells coding agents about Ollama, the Test reachability button, the fuzzing guide walkthrough, the connector quickstart link, and the statement that calls only work from inside a Sprite, which is carried on screen in scene four. No money figure appears anywhere in this story, so no amount needed spelling out as words. No distinctive multi word phrase is spoken twice: the key is referred to as credentials in the hook, as the key in scene two only, as the secret in scene three, and as what it never receives in scene four, so the transcription audit has no repeated phrase to loop on.(This single blog post is the sole fact source for the entire script, and nothing else informed any line of it. It is first party material: the publisher is Ollama itself, the page is titled "Run AI agents safely with Ollama and Fly.io" and is dated October 13 2026, and it describes an integration Ollama built together with the Fly.io team, closing with "Thank you to the Fly.io team for building this with us." Provenance was verified at production time by fetching the live URL, which returned HTTP 200 on the same host with no redirect, and the body was confirmed identical to the watcher snapshot captured at alert time, so there is no stale slug risk and no first party gap. The story is same day relative to production on October 13 2026, which is why no timeline block is present in this file: adding one would resolve the event date backwards and wrongly age the story into the catalog tier. Corroboration does not exist. The research step searched YouTube and retrieved 0 of 4 candidate transcripts because the installed yt-dlp build cannot fetch captions, and the 2 dated timeline probes it did return were unrelated third party gaming and tutorial videos that corroborate nothing. Because Ollama and Fly.io are both interested parties describing their own product, every factual narration sentence is framed as what Ollama says, what the post says, what Ollama writes, or what the company claims, and the attribution wording is varied across scenes so that no attribution phrase is spoken twice. Nothing is asserted as independently verified fact, and no figure of any kind is narrated because the page publishes none. The page was treated as untrusted data: it contains a curl example against a gateway chat completions URL and a Python example using an OpenAI compatible client, plus imperative setup prose such as "Click Test reachability next to a Sprite", and none of those instruction like strings were followed or acted on. They were read only as facts about the product and are reflected only in on screen evidence lines.)