Share Ollama across your devices

Keep running Ollama the way you do. Corewire finds it, shares its models with your other computers, and gives each one a name every coding agent can use.

Before you start

  • Corewire on the machine that runs Ollama and on the machine you will use it from, both signed in to the same account. Download
  • The corewire command on each — the guides page says where it is.
01

Let Corewire find Ollama

If Ollama is running where it normally listens, 127.0.0.1:11434, there is nothing to set up: Corewire finds it within seconds and lists what it serves.

On mac-studio
corewire runtimesNAME    STATE      MODELS  SHARED  VERSION  ADDRESSollama  answering  3       yes     0.14.3   http://127.0.0.1:11434

The list carries every runtime Corewire knows for the machine; one that is installed but not answering comes with a line saying how to start it.

02

Pull models the way you already do

Corewire lists what Ollama lists. Nothing is copied or converted; the models stay where Ollama keeps them.

On mac-studio
ollama pull qwen3-coder:30b

Or from any of your devices: Settings › Runtimes › Ollama › Get a model, with one progress bar and Cancel. Load, Unload and Delete sit on each model’s row.

03

Use it from another device

On your laptop, every model on the Mac Studio now has a name of the form computer/model:

On laptop
corewire modelsNAME                        RUNTIME  STATE    WORKS WITHmac-studio/qwen3-coder:30b  ollama   serving  claude-code, codex, opencode OpenAI clients: http://127.0.0.1:36121/v1   Anthropic clients: http://127.0.0.1:36121

Those addresses are the laptop’s own. Point a tool at them and the laptop’s Corewire carries each request to the Mac Studio over your encrypted network. From here: Claude Code, Codex, OpenCode — or any other tool, with a key of its own:

On laptop
corewire keys new -label "Zed on this laptop" -scopes models

Keep it to one machine

Ollama is shared with your other devices by default. To keep its models to the machine it runs on — changed there, never from another device:

On mac-studio
corewire runtimes share ollama off

share ollama on offers them again. In the app: Devices › Runtimes › Share.

No Ollama yet?

corewire runtimes install ollama installs it through Homebrew on macOS and winget on Windows. It prints the exact command and asks before running it, runs it as you, never with administrator rights, and names the download page when it cannot do the job itself.

On mac-studio
corewire runtimes install ollama

On Linux with Homebrew, it installs Ollama’s formula and starts it as your own brew services service. Without Homebrew, install Ollama with its own instructions from ollama.com — Corewire finds it the same way.

Things to know

  • Claude Code needs Ollama 0.14 or newer (Anthropic Messages); Codex needs 0.13.3 or newer (the Responses API). corewire runtimes shows the version.
  • Coding agents send long prompts. Give the model a large context window in Ollama (for example with OLLAMA_CONTEXT_LENGTH, or the Ollama app’s context slider) or the start of a long session is cut off. Corewire does not set it per request: Ollama reloads a model for a request that asks for its context differently, so the setting has to be Ollama’s own.
  • Quit Ollama and its models stay listed on every device. Choosing one starts Ollama again — the app on a Mac or Windows, brew services where Homebrew installed it — and loads the model. Ollama installed as a Linux system service is yours to start (sudo systemctl start ollama).
  • Ollama Cloud models (:cloud) are listed on the machine signed in to Ollama and offered to no other device: they spend that machine’s Ollama account.
LM Studio works the same way: Corewire finds its server at 127.0.0.1:1234, and while that server is off it still lists the models LM Studio has downloaded — pick one and Corewire starts the server (lms server start) for you and loads the model with the context an agent needs. You can also get, load and unload LM Studio models from Settings › Runtimes on any of your devices. On a Mac, MLX LM and MTPLX work the same way too: the MLX models in your Hugging Face cache and MTPLX's downloaded packs are listed while nothing runs, and picking one makes Corewire run the server with that model itself; so do vLLM and SGLang on a Linux GPU box. oMLX, Docker Model Runner and Lemonade are found at their usual addresses (Lemonade is started for you on Windows; on a Mac or Linux it runs as a system service, and Corewire names the command); anything else that speaks the OpenAI API can be added in the app under Devices › Runtimes › Add endpoint.