n8n-nodes-openrouter-advanced

An n8n community node for the OpenRouter API.

n8n ships an OpenRouter chat model node, but it only exposes nine settings. This package covers the inference side of the API properly: chat, responses, Anthropic-style messages, image generation, video generation, text-to-speech, transcription, embeddings and reranking — plus the OpenRouter-specific controls that make it worth using in the first place, like provider routing, plugins, reasoning effort and hard price ceilings.

Two nodes

OpenRouter — a regular action node. Drag it into a workflow to call any of the endpoints below.

OpenRouter Chat Model (Advanced) — an AI sub-node. Connect it to an AI Agent, Basic LLM Chain, or any other node that takes a language model. It does everything the built-in OpenRouter Chat Model does, plus provider routing, plugins, reasoning control and model fallbacks.

Both use the existing OpenRouter credential that ships with n8n, so there is no new credential to set up and nothing to re-enter.

What it covers

| Resource | Operations | |---|---| | Chat | Complete | | Response | Create | | Message | Create (Anthropic format) | | Image | Generate, List Models | | Video | Generate, Generate and Wait, Get Status, Download, List Models | | Audio | Text to Speech, Transcribe | | Embedding | Create, List Models | | Rerank | Rerank | | Model | List, Get, List Endpoints, List Providers |

Provider routing

Decide which upstream provider serves each request: order, allow-list, ignore-list, sort by price / throughput / latency / accuracy, quantization filter, zero data retention, data collection policy, latency and throughput preferences, and max price ceilings per prompt token, completion token, image or whole request. The request fails rather than exceeding them.

Plugins

All ten are supported, each with its own settings: Web Search, Web Fetch, File Parser, Moderation, Response Healing, Context Compression, Auto Router, Pareto Router and Fusion.

Reasoning

Effort from minimal through max, plus how much of the reasoning trace to return.

Everything else

Model fallback chains, structured output (JSON object, JSON schema, grammar, Python), tool calling, prompt caching, the full sampling parameter set, session IDs, tracing metadata and leaderboard attribution headers.

Installing

The package is built and symlinked into n8n's custom-node folder:

cd "/run/media/nothingness/New Volume/Custom N8n Nodes/n8n-nodes-openrouter-advanced"
npm install && npm run build
ln -sfn "$PWD" ~/.n8n/custom/node_modules/n8n-nodes-openrouter-advanced

Restart n8n afterwards so it picks the nodes up.

Note: the package lives on the external drive. If the drive is not mounted, n8n will start without these nodes and any workflow using them will fail to load.

Development

npm run build     # compile TypeScript and copy icons
npm run dev       # recompile on change
npm run lint      # n8n community-node lint rules
npm run lintfix   # autofix what it can

Dependency handling

n8n-workflow, @langchain/core, @langchain/openai and @n8n/ai-utilities are not shipped. n8n adds its own node_modules to NODE_PATH at startup, so an installed copy of this package borrows whatever versions n8n itself is running. They are declared as optional peerDependencies and kept in devDependencies only so the TypeScript build can see them.

That means the published package never goes stale against a newer n8n — there is nothing to re-pin after an upgrade.

For local development via a symlink, the story differs: the devDependencies installed next to the source do shadow n8n's copies at runtime, because NODE_PATH is searched last. So after upgrading n8n, match the devDependency versions to the new install and rebuild:

node -p "require('<n8n-install>/node_modules/n8n-workflow/package.json').version"

A quick way to check what an installed copy would actually load:

NODE_PATH=<n8n-install>/node_modules node -p "require('n8n-workflow/package.json').version"

Notes

• Streaming is off on the action node. n8n returns a complete item per execution, so there is nothing to stream into. The sub-node streams normally through LangChain.
• Video generation is asynchronous. Generate submits the job and returns a job ID. Generate and Wait submits, polls, and downloads the finished file, with a configurable poll interval and timeout. If it times out, the job ID is included in the error so you can pick it up later with Get Status.
• Binary data: generated images, speech and video are written to a binary field you name. Transcription and image/video references read from binary fields on the incoming item.
• Enterprise and account-management endpoints (Workspaces, SCIM, Guardrails, Observability, BYOK, API Keys, Credits, Activity, Analytics, Files, Presets) are deliberately out of scope. They can be added as extra resources later without restructuring anything.

Licence

MIT