3. n8n uses the published Ollama URL¶

One published hostname serves both n8n workflow nodes and the AI Assistant.

3.0 Why¶

n8n is on its own VM on a different home computer, so it has no local model. Do not give it one. Installing Ollama next to n8n means CPU inference on a box with no GPU. Pointing it at the Mac’s UTM address (192.168.64.1) cannot work, because that virt LAN exists only between the Mac and its own guest. A mesh VPN would mean enrolling every machine that ever wants a completion. The published api-key hostname from chapter 2 already solves it: n8n sends HTTPS plus a Bearer secret, and the weights and the Metal GPU stay on the Mac.

n8n uses that one hostname in two shapes. Workflow nodes speak Ollama’s native API at the bare origin; the built-in AI Assistant speaks OpenAI-compatible /v1. Same secret, same GPU, two credentials and two models. The Assistant also needs a sandbox and a search backend running beside n8n. Do not mix the two URLs, do not put the Edgible secret on n8n-hooks, do not set the ollama app to None, and do not run n8n or the sandbox on the Mac UTM guest. The Mac serves Ollama (and the website); n8n is the remote self-hosted caller.

n8n VM (other PC)   Basic LLM Chain node ──► https://ollama.<org>.edgible.com      (+ Bearer)
n8n VM (other PC)   AI Assistant chat  ──►  https://ollama.<org>.edgible.com/v1   (+ Bearer)
                                                        │
Ubuntu guest        Edgible serving agent ──► 127.0.0.1:11434 (socat)
                                                        │
macOS host          Ollama.app (Metal) — weights and GPU stay here

Where you run this: almost everything is on the n8n VM (its browser UI and docker compose in ~/n8n); only ollama show / ollama ps and any missing ollama pull run on the macOS host.

Two use cases¶

	1. Invoke the LLM from a workflow	2. n8n AI Assistant
What it is	A node on the canvas (e.g. Basic LLM Chain + Ollama Chat Model). Cron, webhooks, other nodes can call the model.	n8n’s instance chat: it helps build/edit workflows and can run generated code in a sandbox.
Edgible URL	`https://ollama.<org>.edgible.com`, no `/v1` (native `/api/chat`)	`https://ollama.<org>.edgible.com/v1`, with `/v1`
Also needs	One Ollama credential. Enable Thinking → off on the node. Model expression `{{ 'qwen2.5:7b' }}`.	Self-hosted `n8n-sandbox` + SearXNG on the n8n VM, not Daytona.
Thinking	Must be off (node option). 7B cannot think.	Must be on in the model. n8n always sends thinking; there is no off switch. Tag must list thinking in `ollama show`.
Smoke test	Execute workflow → a sentence. Mac `ollama ps` on GPU.	Hello, then: visit edgible.com and summarise using the product with n8n.
Model	`qwen2.5:7b`, proven.	`gpt-oss:20b`, proven for real Assistant chat including search, not only Hello. Not `qwen2.5:7b`.

3.1 The job¶

Do both use cases. Same Edgible secret, two models.

Done when

Assistant: ollama show lists thinking; /v1; gpt-oss:20b; Hello replies; edgible.com + n8n (or similar) uses the web. Workflow: Base URL without /v1; {{ 'qwen2.5:7b' }}; thinking off. Same Edgible secret.
Workflow model is not grey llama3.2, not qwen3-coder. Assistant model is not a non-thinking 7B.
Sandbox on the n8n VM: healthz OK; AI settings test OK; no Edgible app on 8080.
SearXNG URL http://searxng:8080 in AI settings; no public SearXNG port.
Execute returns text; Assistant Hello and a search turn return text; Mac ollama ps is GPU.
ollama app is still api-key. The Mac guest still is not running n8n.

Need first: n8n editor on the n8n VM. Chapter 2. Docker Compose v2. ~4 GB / 2 vCPUs spare on that VM for the sandbox runner. For the AI Assistant: a tag with thinking in ollama show (this chapter uses gpt-oss:20b).

Not this chapter: n8n on the Mac guest, qwen3-coder, Open WebUI/sandbox on the 4 GB UTM VM, Daytona, OpenClaw.

Official sandbox docs: Docker Compose and AI Assistant.

3.2 Two models — the Assistant must support thinking¶

Ollama’s n8n page uses qwen3-coder as an example. Do not pull it (~19 GB).

The n8n AI Assistant always sends thinking. That page has no Think-off control. Compose N8N_INSTANCE_AI_THINKING_ENABLED=false does not stop it. A connection test only hits /v1/models, so it does not prove the tag can think. Hello is the thinking check. A question that needs current web results is the SearXNG check.

On the Mac:

ollama show gpt-oss:20b

You want thinking under capabilities. If that word is missing, do not use the tag for the AI Assistant.

n8n feature	Tag	Thinking	Why
Workflow	`qwen2.5:7b`	Off (node option)	Fast 7B. Cannot think, so turn it off on the node.
AI Assistant chat	`gpt-oss:20b`	Required (model capability)	n8n sends thinking. This tag can (~13 GB). Proven for Assistant work including web search via SearXNG.

Cannot use for the AI Assistant (no thinking): qwen2.5:7b, llama3.1:8b, llama3.2:*, mistral:7b, phi*, deepseek-coder:*, codellama:*. Those are fine for workflows with thinking off.

Can use for the AI Assistant if ollama show lists thinking: gpt-oss:20b (this chapter), qwen3.5:27b (heavier, ~17 GB).

If gpt-oss:20b is missing: ollama pull gpt-oss:20b on the Mac, not in the UTM guest. Then ollama show until you see thinking.

3.3 Two credentials¶

AI Assistant: Settings, self-hosted Ollama. Pick the model before you chat:

On the Mac, confirm thinking: ollama show gpt-oss:20b lists thinking. If you pick qwen2.5:7b here, Hello fails with “doesn’t support thinking” even if the connection test passed.
Endpoint = https://ollama.YOUR-ORG.edgible.com/v1 (with /v1). No ?think=; that does nothing.
API Key = the secret from 2.5.
Model = gpt-oss:20b (same string as ollama ls). Not a 7B. Not any tag whose ollama show lacks thinking.
Save. Connection test succeeds (this still does not prove thinking).

Ollama Chat Model in a workflow needs a separate Ollama credential (do not reuse the Assistant URL):

Base URL = https://ollama.YOUR-ORG.edgible.com (no /v1, no :11434).
Same secret.
Save. Connection test succeeds.

Do not use localhost:11434 or UTM 192.168.64.1 from this VM.

Leave sandbox until 3.4. Workflow model id is §3.5 ({{ 'qwen2.5:7b' }}). Assistant Hello is §3.6.

On the workflow node only: Options → Enable Thinking → off. Do not try to turn Assistant thinking off with env or a URL query.

3.4 Self-hosted sandbox (n8n VM)¶

Do this on the n8n VM only. Do not add these containers on the Mac UTM guest.

The sandbox is not an Edgible app. n8n reaches sandbox-api on the Compose network. Do not publish 8080 / 9090 / 9091. sandbox-runner-1 is privileged Docker-in-Docker, so treat it as root on that VM.

Windows host: Compose inside the Linux VM / WSL2, project under ~/n8n (not /mnt/c/...).

This replaces ~/n8n/docker-compose.yml. The named volume n8n_data keeps existing workflows if you already used that name. Sandbox services match n8n’s Docker Compose install.

mkdir -p ~/n8n
cd ~/n8n

.env has four random strings. SANDBOX_API_KEYS and N8N_INSTANCE_AI_SANDBOX_API_KEY must be identical:

A=$(openssl rand -hex 24)
B=$(openssl rand -hex 24)
C=$(openssl rand -hex 24)
D=$(openssl rand -hex 24)
cat > .env << EOF
SANDBOX_API_KEYS=${A}
SANDBOX_API_RUNNER_REGISTRATION_TOKEN=${B}
SANDBOX_API_RUNNER_API_KEY=${C}
N8N_INSTANCE_AI_SANDBOX_API_KEY=${A}
SEARXNG_SECRET=${D}
EOF

Do not commit .env. Save this as ~/n8n/searxng-settings.yml (stock SearXNG only serves HTML; n8n needs JSON):

use_default_settings: true
search:
  formats:
    - html
    - json

Save this as ~/n8n/docker-compose.yml (whole file):

volumes:
  n8n_data:
  sandbox-tls:

services:
  sandbox-certs:
    image: ghcr.io/n8n-io/n8n-sandbox-service-api:latest
    user: "0:0"
    entrypoint: ["sh", "-c"]
    command:
      - >
        bootstrap-mtls.sh --out-dir /tls --api-san sandbox-api
        --control-san-prefix sandbox-runner &&
        chown -R sandbox-api:sandbox-api /tls/api
    environment:
      NUM_RUNNERS: "1"
    volumes:
      - sandbox-tls:/tls

  sandbox-api:
    image: ghcr.io/n8n-io/n8n-sandbox-service-api:latest
    depends_on:
      sandbox-certs:
        condition: service_completed_successfully
    environment:
      SANDBOX_API_KEYS: ${SANDBOX_API_KEYS}
      SANDBOX_API_RUNNER_REGISTRATION_TOKEN: ${SANDBOX_API_RUNNER_REGISTRATION_TOKEN}
      SANDBOX_API_RUNNER_API_KEY: ${SANDBOX_API_RUNNER_API_KEY}
      SANDBOX_API_GRPC_TLS_CERT_FILE: /tls/api/grpc-server.crt
      SANDBOX_API_GRPC_TLS_KEY_FILE: /tls/api/grpc-server.key
      SANDBOX_API_GRPC_TLS_CLIENT_CA_FILE: /tls/api/ca.crt
      SANDBOX_API_RUNNER_CONTROL_GRPC_TLS_CA_FILE: /tls/api/ca.crt
      SANDBOX_API_RUNNER_CONTROL_GRPC_TLS_CERT_FILE: /tls/api/control-grpc-api-client.crt
      SANDBOX_API_RUNNER_CONTROL_GRPC_TLS_KEY_FILE: /tls/api/control-grpc-api-client.key
      SANDBOX_API_RUNNER_CONTROL_GRPC_TLS_SERVER_NAME: sandbox-runner-1
    volumes:
      - sandbox-tls:/tls:ro
    healthcheck:
      test: ["CMD", "wget", "-qO-", "http://localhost:8080/healthz"]
      interval: 5s
      timeout: 3s
      retries: 5
      start_period: 10s

  sandbox-runner-1:
    image: ghcr.io/n8n-io/n8n-sandbox-service-runner-dind:latest
    privileged: true
    depends_on:
      sandbox-api:
        condition: service_healthy
    environment:
      SANDBOX_RUNNER_API_KEYS: ${SANDBOX_API_RUNNER_API_KEY}
      SANDBOX_RUNNER_REGISTRATION_TOKEN: ${SANDBOX_API_RUNNER_REGISTRATION_TOKEN}
      SANDBOX_RUNNER_API_GRPC_ADDR: sandbox-api:9090
      SANDBOX_RUNNER_HTTP_BASE_URL: http://sandbox-runner-1:8080
      SANDBOX_RUNNER_CONTROL_GRPC_LISTEN_ADDR: ":9091"
      SANDBOX_RUNNER_CONTROL_GRPC_ADVERTISE_ADDR: sandbox-runner-1:9091
      SANDBOX_RUNNER_ID: runner-1
      SANDBOX_RUNNER_DOCKER_SANDBOX_IMAGE: ghcr.io/n8n-io/n8n-sandbox-service-sandbox:latest
      SANDBOX_RUNNER_REGISTRATION_GRPC_CA_FILE: /tls/runner/ca.crt
      SANDBOX_RUNNER_REGISTRATION_GRPC_CERT_FILE: /tls/runner/grpc-client.crt
      SANDBOX_RUNNER_REGISTRATION_GRPC_KEY_FILE: /tls/runner/grpc-client.key
      SANDBOX_RUNNER_REGISTRATION_GRPC_SERVER_NAME: sandbox-api
      SANDBOX_RUNNER_CONTROL_GRPC_TLS_CERT_FILE: /tls/runner/control-grpc-server.crt
      SANDBOX_RUNNER_CONTROL_GRPC_TLS_KEY_FILE: /tls/runner/control-grpc-server.key
      SANDBOX_RUNNER_CONTROL_GRPC_TLS_CLIENT_CA_FILE: /tls/runner/ca.crt
    volumes:
      - sandbox-tls:/tls:ro

  searxng:
    image: ghcr.io/searxng/searxng:latest
    environment:
      SEARXNG_SECRET: ${SEARXNG_SECRET}
    volumes:
      - ./searxng-settings.yml:/etc/searxng/settings.yml:ro

  n8n:
    image: docker.n8n.io/n8nio/n8n
    restart: unless-stopped
    depends_on:
      sandbox-api:
        condition: service_healthy
    ports:
      - "127.0.0.1:5678:5678"
    env_file: .env
    environment:
      N8N_PORT: "5678"
      N8N_PROXY_HOPS: "1"
      GENERIC_TIMEZONE: Australia/Adelaide
      TZ: Australia/Adelaide
      N8N_ENFORCE_SETTINGS_FILE_PERMISSIONS: "true"
      N8N_ENABLED_MODULES: instance-ai
      N8N_INSTANCE_AI_SANDBOX_ENABLED: "true"
      N8N_INSTANCE_AI_SANDBOX_PROVIDER: n8n-sandbox
      N8N_INSTANCE_AI_SANDBOX_API_URL: http://sandbox-api:8080
      N8N_INSTANCE_AI_SANDBOX_IMAGE: ghcr.io/n8n-io/n8n-sandbox-service-sandbox:latest
      N8N_INSTANCE_AI_SEARXNG_URL: http://searxng:8080
    volumes:
      - n8n_data:/home/node/.n8n

${SANDBOX_API_KEYS} and the other ${…} lines are Compose substituting from .env, not shell.

cd ~/n8n
docker compose up -d
docker compose ps
docker compose exec n8n wget -qO- http://sandbox-api:8080/healthz
docker compose logs sandbox-api | grep -i runner

You want sandbox-api healthy, healthz {"status":"ok"}, and a log line that the runner registered. n8n still only binds 127.0.0.1:5678.

Do not set N8N_INSTANCE_AI_MODEL to Anthropic, so that Ollama stays in the UI. Change GENERIC_TIMEZONE / TZ if you need a different zone. If this n8n already has n8n-hooks, keep WEBHOOK_URL (and any N8N_HOST / N8N_PROTOCOL) on the n8n environment block from n8n chapter 3. Do not add N8N_INSTANCE_AI_THINKING_ENABLED=false to “fix” the 7B. It does not stop Assistant thinking; use gpt-oss:20b instead.

Then AI settings → sandbox n8n-sandbox (not Daytona). Service URL = http://sandbox-api:8080. API key = SANDBOX_API_KEYS from .env (grep SANDBOX_API_KEYS .env). Save. You want a successful test.

Web search is part of the AI Assistant, not an extra. Provider SearXNG. URL = http://searxng:8080. No API key. Same Compose network as n8n, not localhost, not Edgible. Leave Brave empty unless you buy that. Do not publish SearXNG’s port or create an Edgible app for it. Without this URL, Assistant cannot look anything up; gpt-oss:20b will still chat from weights only.

If sandbox-api never goes healthy: docker compose logs sandbox-certs then sandbox-api. Runner crash-loop: token mismatch in .env. Sandbox test fails: key/URL mismatch, or n8n not on this Compose file.

If the stack is already up without SearXNG: add SEARXNG_SECRET to .env, write searxng-settings.yml, add the searxng service and N8N_INSTANCE_AI_SEARXNG_URL as above, then docker compose up -d.

3.4.1 Health check¶

On the n8n VM, from ~/n8n. This is what the UI uses: probes from inside the n8n container, plus n8n on loopback.

cd ~/n8n
echo "=== compose ==="
docker compose ps -a
echo
echo "=== from n8n container ==="
docker compose exec -T n8n wget -qO- http://sandbox-api:8080/healthz && echo
docker compose exec -T n8n wget -qO- "http://searxng:8080/search?q=n8n&format=json" | head -c 200 && echo
echo
echo "=== n8n on this VM ==="
curl -sf http://127.0.0.1:5678/healthz && echo
echo
echo "=== runner registered ==="
docker compose logs sandbox-api 2>/dev/null | grep -i runner | tail -3

You want: sandbox-api healthy, n8n running, searxng running, sandbox-certs exited 0 (one-shot). Healthz {"status":"ok"}. SearXNG JSON that includes "results". n8n /healthz 200. A log line that the runner registered.

Optional. Ollama from the same container (replace host and secret):

docker compose exec -T n8n wget -qO- \
  --header="Authorization: Bearer YOUR-EDGIBLE-SECRET" \
  "https://ollama.YOUR-ORG.edgible.com/v1/models" | head -c 400 && echo

You want JSON that includes qwen2.5:7b and gpt-oss:20b. 401 = bad secret. Timeout/HTML = Mac Ollama/forwarder down or wrong host.

Do not publish sandbox or SearXNG ports to prove this. ss on the VM should still show 5678 on 127.0.0.1 only.

3.5 One chain (workflow)¶

This is use case 1. You do not need Chat Trigger. Assistant Hello is §3.6.

Add workflow. Search Manual → Manual Trigger (or keep n8n’s “Execute workflow” start node).
Search LLM Chain → Basic LLM Chain. Connect trigger → chain.
On the chain’s Model stub, Ollama Chat Model. Credential Base URL is the Edgible origin without /v1 (not the Assistant endpoint).
The Model dropdown often stays empty and shows grey llama3.2 (n8n’s placeholder, not your Mac). Open the field fx / Expression and set exactly:

{{ 'qwen2.5:7b' }}

Options → Enable Thinking → off.
Chain prompt: Say hello in one sentence. Save, then Execute workflow.

Smoke test (n8n VM). First run can take several seconds (Mac cold load). You want a short sentence in the chain output and Mac ollama ps on GPU.

3.6 Assistant Hello¶

This is use case 2. Sandbox from 3.4 must already test OK. The Assistant model must support thinking (§3.2), because n8n will send it. gpt-oss:20b over Edgible /v1, with sandbox + SearXNG on the n8n VM, is the proven stack, including questions that need a web search, not only Hello.

AI settings: endpoint with /v1, same Edgible secret, model gpt-oss:20b (ollama show lists thinking). SearXNG URL http://searxng:8080.
Open the Assistant chat. Send Hello. You want a reply, not “doesn’t support thinking”.
Then a question that needs the public site, not weights. This one is proven:

Visit www.edgible.com and summarise how I can use the product with n8n.

Smoke test (n8n VM). You want a summary that matches the site (self-host, publish n8n, and similar), not a generic n8n blurb and not a refusal to look anything up. Similar prompts that name a URL and ask for a product summary are the same test. Mac ollama ps shows gpt-oss:20b on GPU.

First run can take a while (20B cold load). If Hello errors on thinking, the model is wrong, not the URL, the sandbox, or N8N_INSTANCE_AI_THINKING_ENABLED. If Hello works but search does not, fix SearXNG (§3.4), not the model.

Verify¶

Assistant: ollama show lists thinking; /v1; gpt-oss:20b; Hello replies; edgible.com + n8n (or similar) uses the web. Workflow: Base URL without /v1; {{ 'qwen2.5:7b' }}; thinking off. Same Edgible secret.
Workflow model is not grey llama3.2, not qwen3-coder. Assistant model is not a non-thinking 7B.
Sandbox on the n8n VM: healthz OK; AI settings test OK; no Edgible app on 8080.
SearXNG URL http://searxng:8080 in AI settings; no public SearXNG port.
Execute returns text; Assistant Hello and a search turn return text; Mac ollama ps is GPU.
ollama app is still api-key. The Mac guest still is not running n8n.

Next¶

4. OpenClaw uses that URL. Series: README.

Last updated 28 August 2026