Cards / ollama-osx
Ollama on an Apple Silicon Mac, natively on its GPU, as an API behind a key
Stewarded by edgible-orgEdgible cards 31/53
| App | What | Who can reach it | Place |
|---|---|---|---|
ollama | Ollama API on the Mac's GPU, through a small Caddy proxy that lets a key use models but not manage them | API key | ollama |
A place is one serving device. Minimum recommended for each:
| Place | Memory | Disk | Architecture | GPU |
|---|---|---|---|---|
ollama | 8 GB | 2 GB | arm64 | Apple GPU |
The first start downloads the app's images: about 598 MB, a few minutes on a typical home connection. That time is the download from the app's registry, not Edgible, and it happens once: later starts and upgrades reuse the images and fetch only what changed.
Found while testing this card. Each is a request for the app's authors, not a fault in the card.
Host names (opens in a new tab) (not yet reported, found 2026-10-27). For ollama (opens in a new tab).Setting it up by hand? Install Edgible from edgible.com (opens in a new tab), then follow the steps below.
Ollama (opens in a new tab) runs open models, such as Llama, Qwen, Mistral, and gpt-oss, on your own machine, behind one API that many tools speak. On an Apple Silicon Mac it is fast because it runs them on the Mac's GPU. In Docker it cannot: Docker on macOS runs a Linux VM with no access to the GPU, so the same models run several times slower on the CPU. This card runs Ollama on the Mac itself, and Edgible publishes it through a small proxy, so your tools and AI agents can use your models from anywhere, behind an API key, without a cloud provider seeing your prompts, and without anything on your own network being able to reach Ollama. On Linux or Windows, Ollama runs well in Docker, and an ollama card for that is planned.
One app, ollama, the Ollama API on api-key: without a key it answers 401, and with one any client that speaks Ollama's API works, such as Open WebUI, Continue, or your own scripts, with the key sent as Authorization: Bearer <key>. Ollama has no login of its own, so the key is what keeps strangers off your GPU. Edgible checks the key on this Mac, before a request reaches anything, and accepts it as Authorization: Bearer <key> or X-API-Key: <key>.
Why there is a proxy. Ollama refuses any request whose Host is not localhost when it listens on 127.0.0.1 only, and Edgible forwards the app's own hostname, so a published, loopback-only Ollama answers every call with 403, key or not. Listening on every address would fix that, and would let anything on this Mac's network use Ollama without a key. Instead, Ollama stays on 127.0.0.1 (OLLAMA_LOCAL_PORT), and a small Caddy (opens in a new tab) proxy on 127.0.0.1 too (OLLAMA_PORT) sets Host to localhost and passes on only what a client needs to use a model: chat, generate, embeddings, the model list and its details, and the OpenAI-style /v1 routes. Pulling, creating, copying, pushing and deleting a model are refused with 403, so a key can use your models and cannot change them: do those on the Mac, against 127.0.0.1:$OLLAMA_LOCAL_PORT. Edgible publishes the proxy, ollama, and puts the key in front of it.
The models live in the folder OLLAMA_MODELS names, by default models/ in the card's directory. Plan for them: a small model takes 2 to 5 GB, a large one 20 to 40 GB, and a model needs about its own size in memory to run. Minimum recommended for the place ollama: an Apple Silicon Mac with 8 GB of memory and 2 GB of disk; it runs on the Mac's GPU.
This card runs natively on macOS, not in Docker. install.sh (opens in a new tab) installs Ollama 0.40.4, checked against its SHA-256, and Caddy 2.11.7, checked against its SHA-512, into the card's own directory and starts them as two background services; uninstall.sh (opens in a new tab) removes both. The settings are card.env (opens in a new tab), the card card.yml (opens in a new tab), and the last test test.yml (opens in a new tab).
Five steps. Edit card.env (opens in a new tab) before you start.
mkdir -p ollama-osx
curl -fsSL https://github.com/edgible-org/cards/archive/refs/heads/main.tar.gz \
| tar -xz --strip-components=2 -C ollama-osx cards-main/ollama-osx
Set DEVICE and ORG_LABEL. Keep OLLAMA_LOCAL_PORT at 11434, the port Ollama's clients expect, unless another Ollama already runs on this Mac: Install stops and names the program that holds a port, and then you set OLLAMA_LOCAL_PORT to a free one. OLLAMA_PORT is the port Edgible publishes; change it only if Check or Install says it is taken. To keep the models an earlier Ollama install downloaded, set OLLAMA_MODELS=~/.ollama/models, and stop that Ollama first (quit the Ollama app, or brew services stop ollama), so two never use the folder at once.
nano ollama-osx/card.env
curl -fsSLo check-env.py https://raw.githubusercontent.com/edgible-org/card-kit/main/check-env.py
python3 check-env.py ollama-osx
It checks this machine, the port, and the service name, and prints any fix as lines to paste. Run it again until it says no conflicts.
set -a
. ollama-osx/card.env
set +a
sh ollama-osx/install.sh
It prints the version, where the models are, and the proxy's port. Both services start again at each login, and run while your user is logged in: for a Mac that serves while nobody uses it, turn on automatic login. Then download a model, such as curl http://127.0.0.1:$OLLAMA_LOCAL_PORT/api/pull -d '{"model": "llama3.2"}', or with the release's own command: OLLAMA_HOST=127.0.0.1:$OLLAMA_LOCAL_PORT ollama-osx/.native/ollama-*/ollama pull llama3.2.
set -euo pipefail
set -a
. ollama-osx/card.env
set +a
device_id=$(edgible device list --json | jq -er --arg name "$DEVICE" '
map(select(.name == $name))
| if length == 1 then .[0].id
else error("need exactly one device named " + $name + " (" + (map(.status + " " + .id) | join(", ")) + ")")
end
')
edgible app create existing \
--non-interactive \
--name ollama \
--port "$OLLAMA_PORT" \
--protocol https \
--auth-modes api-key \
--device-id "$device_id"
edgible app list
ollama uses api-key. Without a key, it answers 401.
curl -sS -o /dev/null -w '%{http_code}\n' "https://<ollama hostname>"
Make a key for each person or tool, which is shown once: edgible app api-keys create --app ollama --name <who>. Then curl -H "Authorization: Bearer <key>" https://<ollama hostname>/api/tags lists your models. edgible app api-keys list --app ollama and delete manage the keys. The card's tests also ask a tiny model a question and check, with /api/ps, that it ran 100% on the GPU; make a key and call the published app with it (401 without the key, 200 with it, a streamed answer, and 403 for pull, create, copy, delete and push), then delete the key; and check that Ollama and the proxy listen on 127.0.0.1 only.
The rest of the setup is in the Ollama docs (opens in a new tab).
Three steps, in this order. Step 1 runs wherever edgible is logged in, steps 2 and 3 on the machine that runs the card.
edgible app delete ollama --yes
set -a
. ollama-osx/card.env
set +a
sh ollama-osx/uninstall.sh
It stops both services and removes them, Ollama and the proxy. The models stay: models/ in the card's directory goes in step 3, and a folder elsewhere (OLLAMA_MODELS) is never deleted, because it is not this card's.
rm -rf ollama-osx