This isn’t a launch recap. Ollama is a way to run a model on your own machine and talk to it over HTTP, including an OpenAI-compatible endpoint that a coding tool can point at. I didn’t install Ollama for this post, and I didn’t run the commands below. What follows is what the current official docs say.

What I verified today
I checked the public GitHub API on November 19, 2026, at 1:14 p.m. America/Phoenix. ollama/ollama was public, not archived, MIT-licensed, and had 187,649 stars. The latest release was v0.40.16, published November 9, 2026, at 11:21 a.m. America/Phoenix. The default branch had last been pushed November 18, 2026, at 3:53 p.m. America/Phoenix. The README on main and at the v0.40.16 tag was byte-identical. The official docs site at docs.ollama.com/quickstart and the OpenAI compatibility page both returned HTTP 200.
Install, the way the docs spell it
The README’s download section gives these install commands. On macOS or Linux:
curl -fsSL https://ollama.com/install.sh | sh
On Windows:
irm https://ollama.com/install.ps1 | iex
The README also links manual downloads, and the Linux docs spell out a manual archive install plus ollama serve. Piping a remote script into a shell means you’re trusting that installer. If you’d rather not, use the manual path the same docs describe.
The quickstart says to download Ollama, then open the app or start from a terminal with:
ollama
“Follow the setup prompts,” the page says. “Sign in to use cloud models, or choose a local model.” Those are two different paths, and the docs treat them that way.
Pull a local model and call the API
The official local examples use Gemma 4 E2B, listed in Ollama’s library as gemma4:e2b. That’s Gemma, the open-weight model family in the library — not Gemini 4. The quickstart says to run it on your computer, with no API key required, and notes that the model download is about 7.2 GB. It recommends 8 GB of available VRAM, or unified memory on a Mac, and says that with less VRAM, Ollama can use system RAM, but responses may be slower. That’s Ollama’s guidance, not a measurement I made.
On Linux, the quickstart says to start the server with ollama serve if it isn’t already running. Then download the model:
ollama pull gemma4:e2b
The same page also shows a chat-in-the-terminal path:
ollama run gemma4:e2b
Ollama “downloads the model and starts a chat on your computer,” the docs say. Type /bye to leave.
To call the OpenAI-compatible local endpoint, the quickstart’s local tab uses this request, with no authorization header:
curl http://localhost:11434/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "gemma4:e2b",
"messages": [
{
"role": "user",
"content": "Say hello in one sentence."
}
]
}'
Read the answer from choices[0].message.content, the page says. The same local tab also shows Ollama’s own /api/chat path at http://localhost:11434/api/chat. The docs’ base-URL table lists the local OpenAI-compatible root as http://localhost:11434/v1, next to the hosted cloud root https://ollama.com/v1.
Cloud is a different official path
Don’t flatten this into “it always stays on your machine.” The quickstart is explicit: “Use cloud models with an API key, or run models locally without one. Cloud requests do not require an Ollama installation.” Direct cloud calls go to https://ollama.com/api or https://ollama.com/v1 with Authorization: Bearer $OLLAMA_API_KEY. The authentication page says to create a key at ollama.com/settings/keys, set it in the terminal, and “Keep keys outside browser code and source control.” API keys “do not expire. Revoke a key in API keys settings.”
If you want only local models, the FAQ says you can disable cloud features by setting disable_ollama_cloud to true in ~/.ollama/server.json, or with the environment variable OLLAMA_NO_CLOUD=1, then restarting Ollama. The logs should then show Ollama cloud disabled: true.
Tokens, bind addresses, and what this is not
The authentication page is blunt about the local server: “The local API at http://localhost:11434 does not require authentication.” The FAQ says Ollama “binds 127.0.0.1 port 11434 by default.” You can change that bind address with OLLAMA_HOST. The same FAQ’s configuration examples use OLLAMA_HOST=0.0.0.0:11434 as the way to listen on every interface. I’m not recommending that. If you do expose the port, the official docs still describe that local API as unauthenticated, so anyone who can reach it can send requests.
I’m also not treating this as a sandbox for coding agents. Ollama’s docs describe a model server and an HTTP API, not an isolated environment that confines what a coding tool can do with your files or your shell. Review the changes a tool wants to make the way you’d review a teammate’s.
On data, Ollama’s FAQ says: “Ollama runs locally. We don’t see your prompts or data when you run locally.” That’s their claim. The same answer says that when you use cloud-hosted models, they process prompts and responses to provide the service but “do not store or log that content and never train on it.” I’m not stretching that into a guarantee about every client, plugin, or tunneling setup you might put in front of the server. The FAQ also documents exposing the server on a network and through proxies; I’m leaving those recipes alone.
Pointing a coding tool at the local endpoint
The official quickstart has a “Use a coding agent” section, and the README lists first-party launch integrations. I’m not walking through those CLIs here. If a tool you already use can set an OpenAI-compatible base URL, the docs say the local root is http://localhost:11434/v1. The OpenAI compatibility page shows client setup against that URL and notes that, for local calls, the client’s API key field is “required but ignored.”
That’s the useful setup, for me, in the official pages: a model you pulled yourself, a server that defaults to localhost, and an endpoint a coding tool can call — with the cloud path, the missing local authentication, and the bind-address switch kept in view.
Hero photo: Old millstone, LSU Rural Life Museum – 2 by abbeyprivate, via Wikimedia Commons; public domain; cropped and resized.