Connect ChatGPT to n8n Workflows: Open Source Step-by-Step Guide

Over 900 million people use ChatGPT weekly as of February 2026, yet most teams still copy-paste prompts manually instead of wiring AI into automated workflows. n8n, the fair-code automation platform founded in 2019 and valued at $5.2 billion after SAP's May 2026 investment, solves this with native OpenAI and LangChain nodes that run on your own infrastructure. This guide shows exactly how to connect ChatGPT to n8n using only open source tools — no cloud lock-in, no per-execution fees, full data control.

Quick Answer: Install n8n self-hosted via Docker, add your OpenAI API key to n8n credentials, then use the built-in OpenAI node or LangChain nodes to call ChatGPT models (gpt-4o, gpt-4o-mini, o1) directly in workflows. Trigger via webhook, schedule, or another app — all running on your server with zero vendor markup.

Why Connect ChatGPT to n8n Instead of Using Zapier or Make

Data Stays on Your Infrastructure

n8n's fair-code license lets you self-host on any server — VPS, Kubernetes, even a Raspberry Pi. Unlike Zapier (SOC 2 but multi-tenant) or Make (EU-hosted but closed), your prompts, completions, and customer data never leave your network. A 2025 n8n case study showed a fintech firm processing 2 million monthly AI tasks on a $40/month Hetzner box, saving $18,000 versus Zapier's team plan.

Native LangChain Integration Since 2024

n8n added dedicated LangChain nodes in 2024, letting you build retrieval-augmented generation (RAG) chains, agent loops, and multi-step reasoning without writing glue code. LangChain, launched October 2022 and backed by Sequoia, provides the abstraction layer; n8n provides the visual orchestration. Together they replace custom Python scripts that break when OpenAI changes endpoints.

Cost Scales With Usage, Not Seats

OpenAI API pricing (gpt-4o-mini at $0.15/1M input tokens, $0.60/1M output as of 2025) applies directly. No per-user fees, no task limits. A marketing agency running 500 content-generation workflows daily pays ~$12/month in API calls versus $299/month for Make's equivalent tier.

Prerequisites: What You Need Before Starting

OpenAI API Account and Key

Create an account at platform.openai.com, add billing (prepaid credits work), and generate a secret key starting with sk-. Save it — you won't see it again. For production, create a separate key per environment (dev/staging/prod) and set usage limits: $50/month default prevents runaway loops.

Server to Host n8n

Minimum: 2 GB RAM, 1 vCPU, 20 GB disk. Recommended: 4 GB RAM, 2 vCPU for concurrent workflows. DigitalOcean Droplet ($24/mo), Hetzner CX32 ($16/mo), or any VPS with Docker support works. For production, add a reverse proxy (Nginx/Traefik) with Let's Encrypt TLS and basic auth.

Docker and Docker Compose Installed

n8n's official deployment method. Verify with docker --version and docker compose version. On Ubuntu 22.04+: sudo apt update && sudo apt install docker.io docker-compose-plugin. The n8n image pulls from docker.n8n.io/n8nio/n8n (Docker Hub mirror available).

Step-by-Step: Self-Host n8n and Add ChatGPT Credentials

1. Create docker-compose.yml

version: '3.8'
services:
  n8n:
    image: docker.n8n.io/n8nio/n8n:latest
    restart: unless-stopped
    ports:
      - "5678:5678"
    environment:
      - N8N_HOST=your-domain.com
      - N8N_PORT=5678
      - N8N_PROTOCOL=https
      - WEBHOOK_URL=https://your-domain.com/
      - GENERIC_TIMEZONE=America/New_York
    volumes:
      - n8n_data:/home/node/.n8n
volumes:
  n8n_data:

Replace your-domain.com with your hostname or IP. Run docker compose up -d. Access https://your-domain.com:5678, complete the owner account setup.

2. Add OpenAI Credentials in n8n

In n8n UI: Credentials → New Credential → Search "OpenAI" → Select "OpenAI API" → Paste your sk- key → Save. Name it "OpenAI Production" for clarity. Test with the "Test" button — green check means valid key and reachable API.

3. Verify LangChain Nodes Are Available

Open any workflow, click "+" → Search "LangChain". You should see: LangChain Agent, LangChain Chain, LangChain Chat Model, LangChain Embeddings, LangChain Memory, LangChain Output Parser, LangChain Retriever, LangChain Tool. These ship with n8n core since v1.0 (2024). No extra install needed.

Build Your First ChatGPT Workflow: Webhook → AI → Response

1. Create Workflow and Add Webhook Trigger

Workflows → New → Click "Start" node → Search "Webhook" → Select "Webhook" → Path: chat → HTTP Method: POST → Response Mode: "Response Node" (critical for synchronous reply). Save.

2. Add OpenAI Node for Chat Completion

Click "+" after Webhook → Search "OpenAI" → Select "OpenAI" → Resource: "Chat" → Operation: "Message a Model" → Credentials: "OpenAI Production" → Model: gpt-4o-mini (cheapest capable model) → Messages: Add Item → Role: "User", Content: {{ $json.body.message }} → Options → Temperature: 0.7, Max Tokens: 1000.

3. Add Respond to Webhook Node

Click "+" after OpenAI → Search "Respond to Webhook" → Select it → Status Code: 200 → Response Data: {{ $json.choices[0].message.content }} → Response Format: JSON. Connect Webhook → OpenAI → Respond to Webhook. Click "Execute Workflow" → Test via curl:

curl -X POST https://your-domain.com/webhook/chat \
  -H "Content-Type: application/json" \
  -d '{"message": "Summarize n8n in one sentence"}'

Returns JSON with ChatGPT's reply in ~800ms.

Advanced: RAG Pipeline With LangChain Nodes and Local Vector Store

1. Prepare Documents and Embeddings

Create a workflow: Manual Trigger → Read Binary File (PDF/Markdown) → LangChain Text Splitter (chunk size 1000, overlap 200) → LangChain Embeddings (OpenAI, model: text-embedding-3-small) → LangChain Vector Store (Qdrant or Chroma local). Qdrant runs in Docker: docker run -d -p 6333:6333 qdrant/qdrant. Store points in collection "docs".

2. Build Retrieval Chain

New workflow: Webhook (path: ask) → LangChain Retriever (Qdrant, collection: "docs", top K: 4) → LangChain Chain (type: "RetrievalQA", Chat Model: OpenAI gpt-4o-mini, Prompt: custom RAG template) → Respond to Webhook. The chain receives {{ $json.body.question }}, retrieves relevant chunks, passes to LLM with context, returns grounded answer.

3. Example: Internal Knowledge Bot

A 12-person dev team at a SaaS company indexed 300 Notion pages (exported to Markdown) into Qdrant via n8n. Their Slack bot (separate workflow, Slack Trigger → HTTP Request to /webhook/ask) answers "How do I reset staging DB?" with exact runbook steps, cutting repeat questions by 73% in month one.

Comparison: Open Source n8n vs. Hosted Alternatives for ChatGPT Integration

Choosing the right platform depends on data sensitivity, scale, and engineering capacity. The table below compares n8n self-hosted against the two most common alternatives using current 2025 pricing and features.

All three support OpenAI API calls, but only n8n gives you full infrastructure control and LangChain nodes without per-execution markup.

Feature	n8n Self-Hosted	Zapier	Make
Monthly base cost (2025)	$16-40 (VPS only)	$299 (Team)	$299 (Teams)
OpenAI API cost	Direct, no markup	Direct + Zapier tasks	Direct + Make operations
LangChain nodes	Native (40+ nodes)	Via Code step only	Via HTTP module only
Data residency	Your server, any region	US/EU (multi-tenant)	EU (AWS Frankfurt)
Concurrent workflows	Unlimited (hardware bound)	Limited by plan	Limited by plan
Custom code support	JS/Python nodes, npm libs	Code step (limited)	Custom functions

Common Mistakes and How to Fix Them

Mistake: Hardcoding API Keys in Workflow JSON

Why It Hurts: Exported workflows leak secrets to Git repos; rotation requires editing every workflow. Fix: Always use n8n Credentials store. Reference via {{ $credentials.openaiApiKey }} only in node config, never in expressions.

Mistake: Using gpt-4o for Every Task

Why It Hurts: gpt-4o costs 33x more than gpt-4o-mini ($5 vs $0.15 per 1M input tokens). Classification, extraction, and formatting tasks rarely need flagship intelligence. Fix: Route by task: gpt-4o-mini for classification/extraction, gpt-4o for reasoning/creative, o1 for multi-step planning. Add a Switch node keyed on {{ $json.taskType }}.

Mistake: No Retry or Fallback on API Errors

Why It Hurts: OpenAI returns 429 (rate limit), 500, 503 regularly. Unhandled, these crash workflows and lose webhook payloads. Fix: On every OpenAI/LangChain node: Settings → Error Workflow → Create "Error Handler" workflow that logs to Postgres/Slack and retries with exponential backoff (Wait node: 5s, 30s, 120s). Enable "Continue On Fail" for non-critical branches.

Mistake: Skipping Input Validation Before LLM Calls

Why It Hurts: Empty or 50KB prompts waste tokens and trigger model refusals. Fix: Add IF node before OpenAI: {{ $json.body.message && $json.body.message.length > 0 && $json.body.message.length < 8000 }}. False branch → Respond to Webhook with 400 and validation message.

Pro Tips

• Stream responses for UX: OpenAI node → Options → Stream: true → Respond to Webhook with "Response Stream" mode. Cuts perceived latency from 3s to first token in 300ms.
• Cache embeddings: Store text-embedding-3-small vectors in Qdrant with payload {source, hash}. Before re-embedding, query by hash — saves 90% embedding costs on unchanged docs.
• Use n8n's built-in queue mode: Set EXECUTIONS_MODE=queue and run Redis + worker containers. Handles burst traffic (1000+ webhooks/min) without dropping requests.
• Version workflows with Git: n8n CLI n8n export:workflow --all --output=workflows/ → commit to private repo. Rollback = n8n import:workflow --input=workflows/.
• Monitor with LangSmith (free tier): Add LangChain Callback Handler node → LangSmith API key → traces every LLM call, token count, latency. Catches prompt regressions before users complain.

FAQ

What is the difference between n8n's OpenAI node and LangChain Chat Model node?

The OpenAI node calls the Chat Completions API directly with minimal abstraction — best for single-turn prompts, structured outputs, and function calling. The LangChain Chat Model node wraps the same API inside LangChain's interface, enabling chain composition, memory, and tool use. Use OpenAI node for simple request/response; use LangChain nodes when building agents, RAG, or multi-step reasoning.

Can I use local LLMs like Llama 3 instead of ChatGPT?

Yes. n8n's LangChain Chat Model node supports any OpenAI-compatible endpoint. Run Ollama (docker run -d -p 11434:11434 ollama/ollama), pull llama3.1, then in LangChain Chat Model set Base URL to http://host.docker.internal:11434/v1 and API Key to "ollama" (any string). Latency drops to ~200ms on GPU, zero API cost, but reasoning quality trails gpt-4o-mini on complex tasks.

How do I secure the webhook endpoint for production?

Three layers: (1) Nginx/Traefik reverse proxy with HTTPS and IP allowlist or basic auth. (2) n8n webhook node → Authentication → "Header Auth" with shared secret (e.g., X-Webhook-Secret: {{ $credentials.webhookSecret }}). (3) Validate payload schema in first IF node — reject malformed requests before any LLM call. Rotate secrets quarterly via n8n Credentials API.

What happens when OpenAI deprecates a model I use in workflows?

OpenAI typically gives 6-12 months notice (e.g., gpt-3.5-turbo deprecated Jan 2025, sunset July 2025). n8n does not auto-migrate model names. Set a calendar reminder for deprecation dates. Test new models in a staging workflow using the same prompts — compare output quality and token cost before switching production credentials. Pin model versions explicitly (gpt-4o-mini-2024-07-18) to avoid silent behavior changes.

Will n8n add support for OpenAI's Responses API and Agents SDK?

n8n's 2025 roadmap (public GitHub issues) shows active work on Responses API support — the new stateful API replacing Chat Completions for agent workloads. The LangChain nodes already wrap LangChain's OpenAI integration, which added Responses API support in v0.2.20 (June 2025). Expect native n8n node updates within 2-3 months of OpenAI GA releases. Track github.com/n8n-io/n8n/issues label "openai" for exact timing.

Conclusion

Connecting ChatGPT to n8n with open source tools gives you a production-grade AI automation backbone for the cost of a VPS and API credits. Self-hosted n8n eliminates per-task fees, keeps data on your hardware, and ships with 40+ LangChain nodes that turn prompt engineering into visual workflow design. Start with the webhook → OpenAI → respond pattern (15 minutes), then layer on RAG, agents, and streaming as needs grow. The fair-code license means you own the stack — no vendor can deprecate your workflow or raise prices on your automation logic.

• Self-host n8n on a $16/mo VPS; connect OpenAI API key once in Credentials store
• Use OpenAI node for simple calls, LangChain nodes for RAG/agents/memory
• Route tasks to gpt-4o-mini by default; reserve gpt-4o/o1 for reasoning-heavy steps
• Add retries, validation, and streaming on every production workflow

Sources

• n8n - Wikipedia
• ChatGPT - Wikipedia
• OpenAI - Wikipedia
• LangChain - Wikipedia