n8n adds Gemini 4 Argon support for MCP, Assistant and Agents

Building workflows from other AI tools via MCP

The first route runs through the Model Context Protocol (MCP), which n8n describes as an open standard governed by the Agentic AI Foundation (AAIF). It lets AI tools connect to external applications and data sources. n8n has a built-in MCP server that works with any MCP client, so users can build workflows from the AI tool they already use.

According to n8n, the requirements are access to Gemini 4 Argon in a developer tool and an idea of what to build. Setup takes a couple of minutes: turn on instance-level MCP in the n8n instance, then add the server URL to any MCP-compatible tool, such as a chat app, a coding agent or an IDE. Many tools support one-click setup and sign-in with OAuth or an API key.

n8n gives an example. A developer working on a signup form asks their agent to build a workflow that catches new signups from a webhook, adds them to a CRM and sends a welcome message in Slack. According to n8n, the workflow appears in the workspace a minute later, built and tested. The agent stops short of publishing and asks first, which keeps a human in the loop.

The n8n Assistant: own key or Cloud credits

The n8n Assistant is a chat-based agent that builds, edits, tests and debugs workflows from plain language. With Bring Your Own Key (BYOK), it runs on an AI provider account the user already holds rather than a model chosen for them.

How this works depends on where the n8n instance runs:

• Self-hosted: the user supplies the key and decides where data and model calls go. The key is entered once in the instance settings, and the Assistant then runs on that model. The Assistant has no native Gemini option yet, so Gemini is reached through OpenRouter or Google's OpenAI-compatible endpoint (base URL: https://generativelanguage.googleapis.com/v1beta/openai/).
• n8n Cloud: nothing needs to be set up. The Assistant runs on n8n Assistant credits, and the workflows it builds can use Gemini 4 Argon through Gateway credits.

Agents handle the judgment calls

n8n defines an agent as a model handed a goal and a toolkit that figures out the steps on its own. In n8n, users set up an agent, choose a model such as Gemini 4 Argon for the reasoning, and plug in the tools it may use, which can be any node, API or sub-workflow. Memory can be added so the agent keeps context. The agent then decides which tool to call next, loops, and stops when the task is done.

According to n8n, Gemini 4 Argon is available as a model for Agents and in the standalone Google Gemini node from launch. On n8n Cloud it can also be used through Gateway credits, so no separate provider account or extra billing is needed.

The company argues that rigid, step-by-step workflows break down when input gets messy. An agent absorbs the judgment calls, the surrounding workflow stays predictable, and users still set a hard limit on which tools the agent can reach. Its example is an incoming support email: the agent reads it, pulls the customer record from the CRM, cross-references help docs, then drafts a reply, sets the ticket priority or flags a human in Slack.

What it means for n8n users

For n8n users, the post lays out three entry points to the new model, depending on how they work. Those who prefer building from an IDE or chat tool can use MCP. Self-hosted teams that want control over their data can bring their own key to the Assistant, while Cloud users can rely on credits. Those who want automation to handle ambiguous input can put the model inside an Agent. The post also points to Google's published benchmarks for Gemini 4 Argon, though the figures themselves are not included in the text.