OpenRouter Chat Model — custom JSON options

This bundle gives you the PR #21780 feature ("set OpenRouter chat model options as JSON", incl. reasoning / reasoning.effort) applied on top of the current official node ([email protected]), in two forms.

What the PR actually does

It adds a Define Options dropdown to the OpenRouter Chat Model node:

• Using Fields Below — the original typed fields (temperature, max tokens, …).
• Using JSON — a free-form JSON box whose keys are sent straight to the OpenRouter chat-completions body (so you can set reasoning, reasoning.effort, provider, etc.).

Because the node runs LangChain's ChatOpenAI under the hood, a few options (max_tokens, reasoning, reasoning.effort, max_retries, timeout) are mapped to LangChain constructor fields explicitly — otherwise LangChain overwrites them with undefined. That mapping is preserved here.

PR vs current master — was it "too far apart"?

No. Diffing the PR file against [email protected], the only real divergences are:

1. The release added a createOpenRouterFetch wrapper (fixes empty tool-call args from Anthropic-via-OpenRouter). Additive — kept here.
2. Internal helper imports were moved into @n8n/ai-utilities. The PR still uses the old @utils/* / ../N8nLlmTracing paths.
3. getProxyAgent gained a timeout-options argument.
4. The release added a builderHint to the model field.

None of these touch the PR's logic, so it re-applies cleanly. The version in patch/ even keeps the proxy timeout and the tool-args fetch fix, which the original PR had dropped.

Option A — patch/LmChatOpenRouter.node.ts (recommended, most robust)

The official node, on top of [email protected], with the PR feature merged in. It still imports n8n's internal modules, so it is meant to be dropped into a fork / custom build of n8n, not installed as a community package.

packages/@n8n/nodes-langchain/nodes/llms/LmChatOpenRouter/LmChatOpenRouter.node.ts

Then pnpm install && pnpm build (or rebuild just @n8n/n8n-nodes-langchain) and run your custom image. This route shares LangChain instances with the Agent node, so there are no compatibility surprises.

Built against: @langchain/[email protected], @langchain/[email protected] (n8n 2.22.5 catalog). If you target a different n8n version, check that release's catalog.

Option B — n8n-nodes-openrouter-custom/ (standalone community node)

A self-contained community package — install it via Settings → Community Nodes with no custom build. It adds a node called "OpenRouter Chat Model (Custom)" and an "OpenRouter Custom API" credential (same fields as the built-in one: API key + base URL).

To keep it installable, the n8n-internal helpers were removed/replaced:

| Official dependency | In the community node | |---|---| | N8nLlmTracing callback | dropped — no per-call token-usage display in the UI; the model works fully | | makeN8nLlmFailedAttemptHandler | dropped — LangChain's built-in maxRetries still applies | | getProxyAgent | dropped — re-add only if you run behind an HTTP(S) proxy | | getConnectionHintNoticeField | inlined as a static notice | | internal credential type | local OpenRouterCredential interface | | createOpenRouterFetch (tool-args fix) | kept (self-contained) |

Build & install

cd n8n-nodes-openrouter-custom
npm install
npm run build          # tsc + copies the icon into dist/
npm publish            # or: npm pack, then install the tarball

In n8n: Settings → Community Nodes → Install → n8n-nodes-openrouter-custom.

⚠️ The one caveat that actually matters: LangChain instance matching

An ai_languageModel sub-node returns a LangChain ChatOpenAI. n8n's AI Agent must recognise it as a LangChain model. If the community package bundles its own copy of @langchain/*, you can hit instanceof/duplicate-module issues and the Agent may reject the model.

That's why @langchain/core and @langchain/openai are declared as peerDependencies (not bundled) — the goal is to use n8n's own copies. Match them to your n8n version (2.22.5 → core 1.1.41, openai 1.4.4). If your n8n install does not expose those modules to community nodes, or the Agent refuses the model, fall back to Option A (the patched node in a custom build), which is guaranteed to share instances.

This LangChain-isolation point — not the PR/master drift — is the real reason a community node "might not just work", and is the honest answer to "what else can we do": run the patched official node in a custom n8n image.

Using JSON mode

Set Define Options → Using JSON and supply e.g.:

{
    "max_retries": 2,
    "timeout": 60000,
    "temperature": 1.0,
    "reasoning": { "effort": "high" },
    "response_format": { "type": "text" }
}

Keys map to the OpenRouter chat-completion request.