Automating with n8n

How to drive this server from a self-hosted n8n instance: the AI nodes connect through the OpenAI or Ollama credential with the base URL pointed here, and the HTTP Request node reaches the native document API (OCR, extraction, classification, PDF operations, transcription) that no bare model runner offers. Written for automation builders; no code beyond node settings.

1Self-hosted n8n only#

Everything here assumes n8n runs on hardware that can reach this server: the same machine, the same LAN, or a network you control (self-hosting n8n is its own documented path). n8n Cloud, and cloud automation platforms generally, execute on the vendor's infrastructure and cannot reach a private server without exposing it to the internet, which defeats the reason you run models locally. Keep the runner where the data is.

2Connect the AI nodes#

n8n's AI nodes (AI Agent, the chat model nodes, the embeddings nodes, chains and summarizers) speak the OpenAI and Ollama dialects this server serves natively (API Compatibility). Two credential routes, both stable:

• OpenAI credential. Create an OpenAI credential in n8n, set its Base URL to http://your-server:PORT/v1, and paste an API key minted in Access as the key. Every OpenAI-flavored node (chat model, embeddings, the AI Agent behind them) now runs against your models.
• Ollama credential. Point the Ollama credential's base URL at the server root (no /v1). The Ollama-flavored chat and embeddings nodes work unchanged, including model listing.

The model field in either node names a model from your catalog; a request that names none gets the server's default for the capability. Embeddings nodes feed n8n's vector-store operations exactly as they would against a hosted API; the embedding side of the house is Embeddings and Reranking.

The same two routes serve other self-hosted automation builders. Flowise and Dify both register a custom OpenAI-compatible (or Ollama) model provider: point its base URL here, paste a key from Access, and every rule in this guide, the Docker networking trap included, applies unchanged.

3The Docker networking trap#

The most common failure is not AI configuration, it is networking: n8n runs in a Docker container, the server runs on the host, and localhost inside the container is the container, not the host. Two fixes, both required:

1. Address the host, not localhost. On Docker Desktop (Windows, macOS), use http://host.docker.internal:PORT as the base URL. On Linux, add host.docker.internal:host-gateway to the container's extra_hosts (Compose) and use the same name, or use the host's LAN address directly.
2. Let the server accept the connection. A default installation listens on loopback only, and traffic from a container arrives on a bridge interface, not loopback, so it is refused regardless of the address you dial. Switch the posture to Network in Ports and mint a key first, because a network-reachable server refuses keyless calls. The checklist is Going Live.

When both run in containers on one Compose network, use the server's service name as the host and skip host.docker.internal entirely.

4Beyond chat: the document API#

The chat and embeddings nodes are table stakes; the reason to pair n8n with this server is the native surface under /lmkit/v1, called from the HTTP Request node. Authenticate it with a Header Auth credential sending Authorization: Bearer lmk_..., and reuse that one credential across every node below. The full request and response shapes are live in the API reference.

For binary payloads, send the file from the previous node as form-data (the upload endpoint expects the form field file); for JSON endpoints, pass the returned fileId as input with input_format set to FileIdentifier. That upload-once pattern keeps large documents out of base64 and out of n8n's execution data.

5A worked flow: watch a folder, extract invoices#

The classic recipe, five nodes, no code:

The false branch writes the extracted json straight to your database, spreadsheet, or accounting system with n8n's ordinary nodes. The true branch posts the flagged fields to a review channel (email, chat, a ticket) with the per-field confidence and page coordinates from elements, so a person confirms instead of re-reading. Designing schemas that extract well, and why routing on that one flag is the whole game, is Structured Extraction; the same pattern extends to splitting and classification as the full IDP pipeline.

Two variants worth building next: an inbox classifier (Email Trigger (IMAP), then /lmkit/v1/categorize against your own category list, then route by result), and an archive pipeline (upload, document-to-markdown for the searchable text, pdf-to-pdfa for the archival copy).

6Long work rides the jobs contract#

A 200-page scan or an hour of audio should not hold an n8n execution's HTTP connection open. Add the header Prefer: respond-async to the HTTP Request node: the server answers 202 Accepted immediately with a job_id. Then loop a Wait node (a few seconds) into an HTTP Request polling GET /lmkit/v1/jobs/{job_id}, and exit the loop when status is anything other than processing; the finished result is byte-for-byte what the synchronous call would have returned. Retry discipline in workflow terms: retry 503 and 429 after the Retry-After pause, never retry a 4xx unchanged, and remember results are held for one hour, so poll promptly and store them on your side. The full contract is Errors, Retries, and Jobs.

7Stated plainly#

• Self-hosted n8n only: a cloud runner cannot reach a private server, and should not.
• Two credentials cover the AI nodes (OpenAI with a base URL override, or Ollama); one Header Auth credential covers the entire native API.
• From Docker, dial host.docker.internal, set the server's posture to Network, and mint a key first.
• The document endpoints are the differentiator: upload once, chain by fileId, branch on human_verification_required, and let Prefer: respond-async handle anything long.