If you've been watching the automation space, you already know the headline: Zapier launched Agents, Make shipped Maia, and everyone is suddenly an "AI automation platform." The problem is most of these platforms want to charge you per task execution once your agents start doing anything meaningful.

n8n 2.0, released in January 2026, took a different path. Native LangChain integration, 70+ AI nodes, persistent memory, vector store support — and if you self-host it (which you absolutely should), you pay nothing per execution. I've been running n8n on Proxmox for over a year, and the 2.0 release is the biggest leap the platform has made.

Here's what actually changed, what it can do now that it couldn't before, and how to decide if it's worth migrating from whatever you're running today.

What "Native LangChain Integration" Actually Means

Before n8n 2.0, adding AI to a workflow meant calling an OpenAI HTTP node, parsing the JSON output yourself, and duct-taping it to the next step. It worked, but it wasn't agentic — it was just an API call in a flowchart.

With 2.0, n8n bakes in LangChain as a first-class execution model. That means:

Tool Nodes: Your agent can call other n8n nodes as tools. Need the agent to look something up in a Postgres database, send a Slack message, or hit an internal API? You wire those as tools and the LLM decides when to call them. This is real agent behavior, not just prompt chaining.

Persistent Memory: Agents can maintain context across executions using a memory node backed by Redis, Postgres, or in-memory storage. Useful for things like a support bot that remembers the last three messages in a conversation without you managing session state yourself.

Vector Store Nodes: Native integrations with Pinecone, Qdrant, Weaviate, and Supabase pgvector for RAG workflows. Build a knowledge base from your documents, index it, and let agents query it at runtime — all within the same workflow canvas.

Human-in-the-Loop: New Wait nodes let an agent pause execution, send you a summary of what it's about to do, and resume only after you approve. This is the feature most people don't realize they need until their agent starts deleting things.

The Real-World Test: Three Workflows That Actually Work

1. Blog Content Pipeline (What I'm Running Now)

Trigger: weekly cron → search for trending topics via Brave Search API → pass results to an AI agent with a custom system prompt → agent writes a structured post → publish to Firebase via webhook → request Google Search Console indexing via Chrome extension.

This whole pipeline runs on a single Proxmox LXC. Total cost beyond the hardware: $0/week. The equivalent on Zapier with their per-task billing would run $15–40/month just for the AI steps.

For hardware to run this reliably at home, a small server or NUC does the job well — something like an Intel NUC 13 Pro handles Proxmox, n8n, and Ollama simultaneously without breaking a sweat. If you want to run local LLMs inside your agents (Ollama integration is built in), you'll want a GPU-equipped machine or at minimum 16GB of fast RAM.

2. Email Triage Agent

Pull unread emails via IMAP → AI agent classifies them (urgent/action needed/FYI/junk) → important ones get summarized and sent to a Telegram message → low-priority get auto-labeled in Gmail.

The agent node here uses GPT-4o-mini with a classification prompt. Total latency per email: under 3 seconds. I process about 60–80 emails a day and spend maybe 5 minutes on triage now.

3. Document Q&A Over Internal Files

This is the RAG use case. New PDFs dropped in a watched folder → parsed → chunked → embedded using a local Ollama embedding model → stored in Qdrant (running in another LXC) → Telegram bot lets me ask questions and the agent retrieves relevant chunks before answering.

For local embedding models to run fast enough to be useful, you want NVMe storage — a solid 2TB NVMe drive makes a big difference versus spinning disk for vector search workloads.

n8n vs Zapier vs Make: The Honest 2026 Breakdown

Let me save you 10 hours of comparison reading.

Use Zapier if: You're non-technical, need something running in 20 minutes, and are connecting mainstream SaaS tools (Gmail, Slack, Salesforce). It has 8,000+ integrations and the UX is genuinely good. You'll pay for it — roughly $0.01–0.02 per task at scale — but the setup friction is lowest.

Use Make if: You want Zapier-level polish with more control over data flow and lower pricing. Make's visual canvas is excellent for complex branching logic. Their Maia AI assistant is still in beta and not production-ready, but the core platform is solid. Better price/performance than Zapier for moderate volumes.

Use n8n if: You're a developer or sysadmin comfortable with self-hosting, you want real AI agent capabilities (not just prompt steps), and you have privacy or cost constraints that make cloud automation impractical. The UI is less polished than Zapier but the power ceiling is much higher.

For books on automation architecture and building agent systems, Building LLM Powered Applications by Valentina Alto is worth reading if you want to understand what's happening under the hood. The Art of Automation is more accessible if you're coming from a no-code background.

Setting Up n8n 2.0 on Proxmox in 15 Minutes

If you're already running Proxmox, the fastest path is an LXC container:

# Create an Ubuntu 22.04 LXC, then inside it:
curl -fsSL https://deb.nodesource.com/setup_20.x | sudo -E bash -
sudo apt install -y nodejs
npm install -g n8n
For production, use PM2 to keep it alive
npm install -g pm2
pm2 start n8n
pm2 save
pm2 startup

Set N8N_BASIC_AUTH_ACTIVE=true and configure your reverse proxy (Nginx or Caddy) in front of it. If you want HTTPS without opening firewall ports, Cloudflare Tunnels is the cleanest approach — free, zero port exposure.

For a proper home lab server that can run this stack (n8n + Ollama + Qdrant + whatever else), the Minisforum MS-01 is the current value king — Intel i9, supports up to 96GB RAM, two 2.5GbE ports, and fits on a shelf. I'd pair it with 32GB of DDR5 RAM minimum for a comfortable n8n + local LLM stack.

The AI Node That Changes Everything: Agent Node

The single most important addition in n8n 2.0 is the Agent node. It's not a glorified HTTP call to OpenAI — it's a full ReAct-style agent loop that:

1. Receives a task and available tools 2. Reasons about which tool to call 3. Executes the tool (which can be any other n8n node) 4. Observes the result 5. Decides whether to call another tool or return a final answer

You define the tools, system prompt, memory backend, and output format. You can swap the underlying model (GPT-4o, Claude, Gemini, local Ollama) without changing anything else. This portability is genuinely useful — when OpenAI pricing changes or a better model ships, you update one dropdown.

What's Still Rough

Honesty check: n8n 2.0 isn't perfect.

The error messages when an agent workflow fails are often cryptic. Debugging a multi-step agent that's silently returning empty tool results takes patience. The community forum is helpful but you'll hit edge cases that require reading source code.

The UI is also clearly built for power users, not end users. If you're building workflows for non-technical teammates, you'll need to wrap the webhook endpoints in something friendlier. n8n's form nodes help, but it's not a polished product interface.

And if you're relying on cloud-hosted n8n (n8n.cloud), you're back to paying per execution, which narrows the cost advantage significantly. The self-hosted path is where the value lives.

Verdict

n8n 2.0 is the best platform available for developers who want to build real AI agents without recurring per-execution costs. If you're already self-hosting anything on Proxmox or Docker, adding n8n to your stack is a no-brainer.

The LangChain-native agent loop, vector store integrations, and human-in-the-loop patterns put it ahead of Zapier and Make for technical use cases. Zapier remains better for non-technical users who need fast setup and broad integrations. Make sits in a reasonable middle ground.

If you're serious about building agentic workflows and want full data control, spin up an LXC, install n8n, and spend a weekend building your first real agent pipeline. The productivity compounding from having agents that actually do things — not just trigger webhooks — is substantial.

---

Disclosure: This post contains affiliate links. If you purchase through these links, I may earn a small commission at no extra cost to you.