Ollama

What it is

Ollama is an app and command line that runs open models on your own computer, or in Ollama’s cloud once you sign in. Its README tells you to start building with open models and to access the latest releases locally or in the cloud “with complete privacy”. You pull a model by name, chat with it in a terminal, and call it from code through a REST API on localhost port 11434. The command ollama launch points coding agents such as Claude Code, Codex and OpenCode at its models.

1. You install the app, or run the image in Docker
2. ollama run downloads a model and starts a chat
3. The local server answers on localhost port 11434
4. Your code and coding agents send requests there
5. A cloud model runs on Ollama’s servers if you sign in

Who it suits

Developers and tinkerers who want open models on their own machine, and anyone who wants Claude Code, Codex or OpenCode to run on a different model. Skip it if disk or memory is tight. The Quickstart’s small example model is a 7.2 GB download. It recommends 8 GB of VRAM, and the macOS and Windows pages say models can reach tens to hundreds of GB.

Where to use it

Ollama runs on macOS, Windows and Linux, and as a Docker image. Its README and docs name these ways to connect other tools.

• Claude Code, Codex, Copilot CLI, OpenCode and Droid: ollama launch claude, ollama launch codex, ollama launch copilot, ollama launch opencode and ollama launch droid
• DeepSeek’s open-source coding agent, a developer preview: ollama launch dsh
• OpenClaw, a personal assistant that links WhatsApp, Telegram, Slack, Discord and more
• The Claude and ChatGPT desktop apps on macOS, from the Apps screen in the Ollama app
• VS Code through the Ollama extension, and Zed through its LLM provider settings
• Your own code: the REST API, the Python and JavaScript libraries, and OpenAI and Anthropic compatible addresses

How to set it up

Get it saves a zip of Mindket’s setup notes in your downloads folder, with a page for each platform and tool from Ollama’s README and docs. The main route is the install line or the installer for your system.

1. Install Ollama with the line for your system, or download the installer.
2. Open the app, or run ollama in a terminal and follow the prompts. Sign in to use cloud models, or choose a local model.
3. Chat with a model, for example ollama run gemma4:e2b. Type /bye to leave.
4. To connect a coding agent, run ollama launch with the agent’s short name, such as claude or copilot.

curl -fsSL https://ollama.com/install.sh | sh

irm https://ollama.com/install.ps1 | iex

The first line is for macOS and Linux, the second for Windows PowerShell. Both download a script and run it at once.

ollama run gemma4:e2b

How to use it

Pull and chat from the terminal, or point a tool at the server. A few commands cover daily use.

ollama pull gemma4
ollama ls
ollama ps
ollama stop gemma4
ollama rm gemma4

In order, they download a model, list what you have, show what is loaded, stop a loaded model and delete one. Running ollama launch with no name opens an interactive setup for the tools it supports. Other programs reach the server at http://localhost:11434.

Example

You want Claude Code to answer questions about a project using an Ollama model. From the project folder you run this.

ollama launch claude

The Quickstart says Ollama offers to install Claude Code if it is missing, and you choose a model when prompted. Then you type the Quickstart’s prompt.

Explain how this repository is organized.

For a bigger repository the Claude Code page says to set the context length to 64k or higher. In the app that is a slider under Settings, and from the command line it is OLLAMA_CONTEXT_LENGTH=64000 ollama serve.

For newcomers and experts

If you are new. A model is a large file that predicts text. Ollama downloads it, loads it into memory and gives you a chat prompt plus a local address that other programs can call. Start with ollama run gemma4:e2b and ask it something. Then run ollama ps in a second terminal. Its PROCESSOR column says whether the model sits on the GPU, the CPU or both.

For experts. The README names llama.cpp as its supported backend. Your local server at http://localhost:11434 and ollama.com serve the same APIs: the Ollama API, OpenAI Chat Completions and Responses, and Anthropic Messages. Changing the base address, and adding an API key for ollama.com, moves a tool between local and cloud. The docs say the OpenAI and Anthropic routes each cover a subset of the original API, and Responses requests are stateless.

The FAQ says required memory scales with OLLAMA_NUM_PARALLEL, which defaults to 1, times OLLAMA_CONTEXT_LENGTH. OLLAMA_MAX_LOADED_MODELS defaults to 3 times the number of GPUs, or 3 on CPU, and OLLAMA_MAX_QUEUE to 512. Flash Attention is used when the backend and devices support it. With Flash Attention enabled, OLLAMA_KV_CACHE_TYPE can set the K/V cache to q8_0, which the FAQ says takes about half the memory of f16, or q4_0 at about a quarter.

ollama launch leaves OpenCode’s own config file alone, according to the OpenCode page. It gives Codex a dedicated profile, which ollama launch codex --restore removes.

Good to know

Most of it runs on your machine, and a few paths reach out.

• Accounts: local models need no account. Cloud models need a sign-in or an API key, and the Cloud page links your usage (ollama.com/settings/usage) and its plans (ollama.com/pricing). The Quickstart points to it for usage limits
• What leaves your machine: the FAQ says the makers do not see your prompts when you run locally. For cloud models it says prompts and responses are processed but not stored or logged, and never used for training
• Account data: the FAQ says Ollama collects basic account info and limited usage metadata for cloud, and that it does not include prompt or response content
• Cloud off: OLLAMA_NO_CLOUD=1, or disable_ollama_cloud in ~/.ollama/server.json, turns off cloud models and web search
• The local server: it binds to 127.0.0.1 port 11434 by default, and the Authentication page says the local API needs no authentication. Think before you set OLLAMA_HOST to expose it
• Install lines: they pipe a downloaded script into your shell, so read the script first if you want to know what it does
• Updates: on macOS and Windows the app downloads updates by itself, and you apply them from the menu bar or taskbar item with Restart to update. It also registers as a login item at install. Task Manager’s Startup apps tab or macOS Login Items stops it starting at login
• Context length: the default follows your VRAM, and the Context length page asks for at least 64000 tokens for agents and coding tools. The FAQ still says 4096, so check yours with ollama ps
• Other tools: ollama launch offers to install Claude Code, OpenCode and OpenClaw, but Codex must be installed first. OpenClaw shows a security notice about tool access on first launch, and DeepSeek’s coding agent is a developer preview whose upstream configuration may change between releases
• Models: they live in a .ollama folder in your home directory on macOS and Windows, and under /usr/share/ollama on Linux. OLLAMA_MODELS moves them, and the docs give removal steps for macOS, Windows and Linux but not Docker
• License: MIT. The LICENSE file reads “Copyright (c) Ollama”