Skip to main content
This tutorial walks through building an n8n workflow where users can upload CSV files through a chat interface and get TabPFN predictions back. The agent parses the file, sends the data to TabPFN via MCP, and returns results in the conversation. Works with both n8n Cloud and self-hosted instances.

​
Prerequisites

​
What you can build with this

Once you have an n8n agent connected to TabPFN, you can wire it into any n8n automation. A few examples:
Use caseHow it works
Lead scoringUpload a CSV of leads, get churn/conversion probabilities, route high-value leads to a CRM node
Outreach gatingRun predictions mid-workflow and branch on the result — continue the sequence or stop
Anomaly flaggingScore incoming data against a baseline and trigger Slack/email alerts on outliers
The workflow in this tutorial handles the CSV upload and prediction part. You can extend it with additional n8n nodes for whatever comes after the prediction.

​
Overview

Here’s the finished workflow:
Complete n8n workflow showing chat trigger, If node, Extract from File, and AI Agent with Anthropic model
The flow:
  1. A chat trigger receives messages and file uploads from the user.
  2. An If node checks whether the message includes a file.
  3. If yes, an Extract from File node parses the CSV into structured data.
  4. The AI Agent — connected to TabPFN via MCP — handles the conversation, receives the parsed data, and runs predictions.

​
Step 1: Chat trigger with file uploads

Create a new workflow and add a When chat message received trigger node. Open the node, click Add Field under Options, and enable Allow File Uploads. You can also add Allowed File Mime Types and set it to text/csv if you want to restrict uploads to CSV files only.
Chat trigger node with Add Field dropdown showing Allow File Uploads option

​
Step 2: Route based on file uploads

Add an If node connected to the chat trigger output. Set the condition to check whether the incoming message has files attached:
  • Field: {{ $json.files }}
  • Condition: is not empty
When a user sends a message with a CSV, the files array in the trigger output will contain the file metadata and binary data. Messages without files go down the false branch.
If node configured to check $json.files is not empty
If you need to handle multiple file types (Excel, JSON, etc.), add a second condition or a Switch node after the If to route each type to its own extraction node.

​
Step 3: Extract CSV data

Connect an Extract from File node to the true branch of the If node. Configure it:
  • Operation: Extract From CSV
  • Input Binary Field: data0
The chat trigger stores uploaded files as binary data. The first file is data0, the second would be data1, and so on. The Extract from File node converts the binary CSV into structured JSON rows that the agent can work with.
Extract from File node set to Extract From CSV with Input Binary Field data0

​
Step 4: Connect the AI Agent

Add an AI Agent node to the workflow. Connect it so it receives data from both branches:
  • The true branch (after Extract from File) feeds in the parsed CSV data.
  • The false branch passes through chat messages that don’t have files.
The AI Agent must also be connected to the chat trigger node. This is required — without this connection, the chat interface won’t return responses. In n8n, the chat trigger needs a direct path to an agent node for the conversation loop to work. Connect your LLM to the agent’s Chat Model input. This tutorial uses the Anthropic Chat Model node, but you can use OpenAI, Google, or any other supported model.
Workflow showing chat trigger connected to If node

​
Step 5: Add conversation memory

Add a Simple Memory node and connect it to the agent’s Memory input. Configure it:
  • Session ID: Connected Chat Trigger Node
  • Session Key From Previous Node: {{ $json.sessionId }}
  • Context Window Length: 15 (or adjust based on how much conversation history the agent should retain)
This keeps the conversation stateful so users can ask follow-up questions about their predictions.
Simple Memory node configured with session ID from chat trigger

​
Step 6: Connect TabPFN via MCP

Add the TabPFN MCP server as a tool for the agent. If you haven’t set up the MCP connection yet, follow the n8n setup guide video first. Connect the MCP tool node to the agent’s Tool input. The agent will have access to TabPFN’s tools for uploading datasets, fitting models, and generating predictions. See the full tool reference for available operations.

​
Step 7: Configure the agent prompt

Set the AI Agent’s system prompt to tell it how to use the TabPFN tools. We maintain a ready-made skill file that covers the full workflow — data validation, tool selection, edge case handling — so you don’t need to write this from scratch. Grab the SKILL.md from the TabPFN Prediction Skill tutorial and paste its contents into the agent’s system prompt field. The skill was written for Claude, but the instructions are model-agnostic — they work with any LLM that supports tool use. You can use the skill as-is for a general-purpose prediction agent, or trim it down for a narrower use case. For a lead scoring workflow, for example, you might hardcode the target column and task type so the agent doesn’t need to ask.

​
Test it

Click Test chat in the n8n editor. Send a message with a CSV file attached — something like: 
Here's my customer data. Predict the "churned" column.
The agent should:
  1. Confirm the target column and task type
  2. Upload the CSV to TabPFN
  3. Return predictions
From here, you can extend the workflow with additional n8n nodes downstream — write predictions to a Google Sheet, send a Slack summary, update a CRM record, or branch into different automation paths based on the prediction results.