n8n workflows with Marina

Start with n8n’s HTTP Request node. It works with Marina’s stateless memory and model APIs, keeps credentials in n8n’s credential store, and makes retries inspectable. Use world MCP when your n8n client can preserve a logged-in MCP session for the whole agent run.

Workflow: retain evidence and a reusable handoff

This recipe needs no model. Start a standalone memory service using memory-service setup. Its provisioned credential file contains a token and spaceId.

1. In n8n, create a Header Auth credential: name Authorization, value Bearer YOUR_MEMORY_TOKEN. This is a memory credential, not a model-provider key.

2. Add a Manual Trigger, followed by Edit Fields. Keep these non-secret fields: memoryUrl, spaceId, eventId, and content. For a first run, use a unique event ID such as demo-handoff-001 and content such as Review the pagination patch before release.

3. Add HTTP Request, named Capture source: POST to {{ $('Edit Fields').item.json.memoryUrl }}/v1/memory/spaces/{{ $('Edit Fields').item.json.spaceId }}/sources. Select Generic Credential Type → Header Auth and your credential. Send JSON, with these fields:

   {
     "content": {"role": "user", "content": "Review the pagination patch before release."},
     "session_id": "n8n-handoff-demo"
   }

   Map content.content from the Edit Fields node for actual input. Add the header Idempotency-Key with expression {{ $('Edit Fields').item.json.eventId + ':source:v1' }}.

4. Add a second HTTP Request, named Remember handoff: POST to the same space’s /records. Use the same credential and the distinct idempotency key {{ $('Edit Fields').item.json.eventId + ':record:v1' }}. Use a JSON expression body:

   {{ {
     content: $('Edit Fields').item.json.content,
     source_ids: [$('Capture source').item.json.id]
   } }}

5. Inspect the returned record ID. Add a GET request to /v1/memory/spaces/SPACE_ID/records/RECORD_ID using that ID and the same credential. The content and source reference should match the input.

Configure memoryUrl for where n8n runs: http://127.0.0.1:3301 only works when it shares the server’s host network. On a container network, use the memory service’s DNS name and a deliberate reachable bind address. n8n Cloud needs a reachable HTTPS deployment. Do not put credentials in Edit Fields, workflow JSON, published panels, or example URLs.

For webhook input, use an immutable upstream delivery/event ID. Reuse the same key and payload after a timeout; a new n8n execution ID on every retry would create a new logical write. Marina rejects reuse of a key with a different payload with HTTP 409. A real edit needs an explicit new operation/version. Capture and remember are two operations: a failed second step may leave a source, and retrying it should reuse the first source receipt. See the memory contract for mutation/version rules.

The n8n HTTP Request reference describes JSON expressions and generic credentials. The credential and node labels may vary with your installed n8n version; the Marina HTTP contract above is the integration boundary.

Ask a Marina model from a workflow

Use an HTTP Request node with POST https://YOUR_MARINA/v1/chat/completions, a separate Header Auth credential containing an actual MODEL_API_KEYS secret, and JSON:

{
  "model": "MODEL_FROM_V1_MODELS",
  "stream": false,
  "messages": [{"role": "user", "content": "Suggest three checks for this release."}]
}

First GET /v1/models with that credential to select a valid model. The reply text is under choices[0].message.content. Model calls can incur provider cost. Keep inference separate from publishing or mutating production systems: a model’s text is not evidence that an action ran. If your n8n OpenAI-compatible model node exposes a configurable base URL, the equivalent base is https://YOUR_MARINA/v1; otherwise use the HTTP Request path.

World participation through MCP

Marina serves Streamable HTTP at /mcp, not a legacy /sse endpoint. n8n’s MCP Client Tool documentation documents bearer/header credentials and tool selection, but does not establish that every version preserves the same session between tools. Verify the node version you deploy:

1. Configure a client transport that supports Streamable HTTP and the actual Marina /mcp URL.
2. Configure the transport bearer if Marina requires it.
3. Call login, then look on the same session. The second call must see the logged-in resident.
4. Reconnect with auth and a saved resident token after transport/session expiry.

If the node only offers legacy SSE or creates a new session per invocation, use the HTTP recipes above or a persistent application bridge such as the LangChain example. Do not assume that an HTTP bearer alone logs in a world resident. Filter tools to the workflow’s needs, and use live capabilities to discover command forms instead of maintaining another catalog.

Let Marina call an n8n workflow

The direction can also be reversed: expose a deliberately selected workflow through n8n’s MCP Server Trigger, then configure it as an outbound Marina MCP connector using MCP connections. Check the node’s production URL and transport. Expose only workflows you intend agents to run; their external actions remain subject to your workflow credentials and approval rules.