Skip to content

5. Tear down the published LLM

The inference hostname gone, the key revoked, and the Mac back on loopback.

5.0 Why

Of everything in these guides, this is the one worth being careful about. ollama is an api-key app in front of a GPU. As long as the hostname exists, anyone holding that secret can run inference on your hardware from anywhere, and a secret you pasted into an n8n credential, an OpenClaw config and a shell history is a secret in more places than you can remember.

The Mac also ends this series in a state it did not start in. Chapter 2 set OLLAMA_HOST to 0.0.0.0:11434, so Ollama is listening on every interface the Mac has, not only loopback. That is fine while the virt LAN is the only network reaching it, and it is worth undoing when you are finished.

Three machines are involved, so this chapter is ordered by blast radius: the public hostname first, then the callers that used it, then the local plumbing, then the weights if you want the disk back.

delete first    app ollama            the public hostname and its api-key
then            n8n credentials       OpenClaw provider config
then            ollama-forward        socat and its systemd unit on the guest
last            OLLAMA_HOST           the Mac back to loopback, weights optional

Where you run this: edgible on the Ubuntu guest, the credential deletes in the n8n UI, openclaw config on the OpenClaw VM, and the last two sections on the macOS host.

5.1 The job

You delete the ollama app, remove the credentials that carried its secret, stop and remove the loopback forwarder, and put the Mac's Ollama back on localhost.

Done when

  • edgible app list no longer shows ollama.
  • The HTTPS origin fails from a phone on cellular, with the Bearer token that used to work.
  • n8n has no Ollama credentials left, and OpenClaw is off ollama/gpt-oss:20b.
  • ollama-forward.service is gone, and nothing is listening on 11434 on the guest.
  • The Mac's Ollama is back on 127.0.0.1:11434.
  • hello-world still loads, and edgible device health is still OK.

Need first: nothing beyond having done the series. Skip any step for a chapter you never did.

Not this chapter: deleting the Ubuntu VM, tearing down n8n on Edgible or OpenClaw on Edgible themselves, or edgible auth logout. Those two guides have their own teardown chapters, and both of those machines can stay exactly as they are.

5.2 Delete the app and the key

On the Ubuntu guest. The app goes first, because deleting it is what makes the secret worthless everywhere it has been copied to:

edgible app list
edgible app delete --name ollama

Deleting the app removes its API keys with it. If you would rather revoke a key while keeping the app, that is edgible app api-keys list --app-id <ollama-app-id> then api-keys delete on the id.

Smoke test. From a phone on cellular, or any laptop off the LAN, with the secret that worked in 2.5:

curl -sS -o /dev/null -w '%{http_code}\n' \
  "https://ollama.YOUR-ORG.edgible.com/api/tags" \
  -H "Authorization: Bearer $EDGIBLE_APP_KEY"

You want a failure: a connection error, or a 404 from the gateway. A 200 with tags JSON means the app is still there. Nothing local has changed yet, which is the point of doing this first: Ollama is still serving on the Mac, and it is no longer reachable from the internet.

5.3 Remove the callers' credentials

The secret is dead, but leaving it in a credential store means it turns up in a backup later, and a workflow pointed at a hostname that no longer resolves fails in a confusing way.

In n8n, from chapter 3, on the n8n VM's browser UI: open Credentials, delete the Ollama credential used by the chat model node and the OpenAI-compatible credential used by the AI Assistant. Then deactivate or delete the workflow that used them. If you replaced ~/n8n/docker-compose.yml in 3.4 to add the sandbox and SearXNG, and you do not want those running any more:

cd ~/n8n && docker compose down

Put back the compose file from n8n on Edgible, chapter 3 before docker compose up -d again, or the editor loses WEBHOOK_URL and your public webhook hostname stops printing the right origin.

In OpenClaw, from chapter 4, on the OpenClaw VM:

openclaw models set google/gemini-2.5-flash
openclaw config delete models.providers.ollama
openclaw gateway restart
openclaw models list --provider ollama

Set whichever model you were on before as the primary. The list should now print nothing for that provider. If config delete is not in your build, empty the values instead: the apiKey is the one that matters.

Smoke test. openclaw agent --model … hello answers on the model you just set, and no chat turn hangs waiting for a hostname that is gone.

5.4 Stop the forwarder on the guest

The socat unit only forwarded loopback to the Mac, and Edgible is no longer publishing that port, so this is tidying rather than exposure. On the Ubuntu guest:

sudo systemctl disable --now ollama-forward.service
sudo rm -f /etc/systemd/system/ollama-forward.service /usr/local/bin/ollama-forward.sh
sudo systemctl daemon-reload

sudo apt-get purge -y socat if nothing else on the guest uses it.

Smoke test. On the guest:

systemctl status ollama-forward.service
ss -ltnp | grep 11434

You want Unit ollama-forward.service could not be found and no output from ss.

5.5 Put the Mac back on loopback

macOS host, Terminal.app. This undoes the bind from 2.2:

launchctl unsetenv OLLAMA_HOST
killall Ollama
open -a Ollama

Smoke test. On the Mac:

lsof -nP -iTCP:11434 -sTCP:LISTEN

You want 127.0.0.1:11434, not *:11434. Local chats still work; nothing on the LAN can reach the model any more.

If you kept a firewall exception for Ollama when you set this up, remove it in System Settings → Network → Firewall → Options.

5.6 The weights, if you want the disk back

Models are the largest thing this series put on your machine: a 7B is roughly 4 GB and gpt-oss:20b roughly 13 GB. They are also the slowest thing to get back, so this is a separate decision from everything above.

ollama ls
ollama rm gpt-oss:20b
ollama rm qwen2.5:7b
ollama rm nomic-embed-text

Keep any tag you use locally. Removing Ollama itself is dragging Ollama.app to the Bin, plus rm -rf ~/.ollama for the model store and history.

Verify

  • edgible app list no longer shows ollama.
  • The HTTPS origin fails from a phone on cellular with the old Bearer token.
  • n8n has no Ollama credentials left, and OpenClaw is off ollama/gpt-oss:20b.
  • systemctl status ollama-forward.service reports no such unit, and ss shows nothing on 11434.
  • lsof on the Mac shows 127.0.0.1:11434, not *:11434.
  • hello-world still loads and edgible device health is OK.

Next

The VM, the serving agent and Hello World are all still there. Website on Edgible publishes a site with self-hosted analytics and monitoring, and n8n on Edgible publishes one process on two hostnames with two auth modes. Series: README.

Last updated