n8n is where your AI agent meets the real world. When the API call fails, the workflow keeps running — and that’s the problem.
n8n is a fair-code workflow automation platform: visual builder, custom JavaScript, and a native AI Agent node — all on one canvas. For SMB deployments, it’s the orchestration layer connecting LLMs to CRMs, databases, Slack, and hundreds of other systems.
What It Does
- Visual workflow builder with branching, loops, and error handling
- Native AI Agent nodes that call tools across your stack
- 400+ integrations plus custom HTTP and code nodes
- Self-hosted (full data control) or n8n Cloud (managed)
Typical Costs
| Deployment | What You Pay |
|---|---|
| Self-hosted | Server only (~$10–25/month VPS); software is free under fair-code license |
| n8n Cloud | Per workflow execution; scales with volume |
| Realistic SMB chatbot (1,000 queries/month) | ~$15–20/month combined hosting + API |
n8n charges by execution (full workflow run), not per step — so a 10-step workflow running 1,000 times costs 1,000 executions, not 10,000.
Where It Breaks
Silent Agent Failures
When a tool inside an AI Agent node fails (400/500 from an API), the Agent node does not fail. The error feeds back into the agent’s “reasoning.” It retries. It burns tokens. No alert fires.
Fix: Wrap tool logic in try/catch, use the HTTP Request node’s “Never Error” toggle, or return structured JSON like {"success": false, "error": "..."} so the agent can self-correct instead of looping.
Maintenance Overhead (Self-Hosted)
Self-hosting means you own uptime, backups, updates, and security patches. For a solo implementer, server maintenance usually eats the subscription savings — unless compliance requires on-premise data.
When to Choose It
- n8n Cloud: You want to ship AI workflows fast without managing servers. The right default for most solo implementers.
- Self-hosted n8n: HIPAA, GDPR data residency, or policies that block third-party workflow execution.
When in doubt, start with Cloud. Migrate to self-hosted only when the compliance case is clear.
Related
- Integration Failure — API drops and masked tool failures
- Silent Agent Failure — Wrong answers with zero errors
- MCP — Protocol for cleaner tool connections
- Integration Layer — Where connections live
- Stack & Tools — Platform profiles
- Zapier — Cloud-only alternative; n8n is the self-hostable counterpart
- Vendor Lock-In — Why n8n’s JSON export matters when you want to leave
- Human-in-the-Loop — The approval step that prevents agent loops from doing damage