n8n Agent Studio

The problem

The Fiverr n8n market is full of demos that use Google Sheets as a database and have no error handling. Serious clients see those and move on. What is missing is an automation portfolio that looks like production: real audit trails, structured tool use, and a single dashboard across workflows.

Goals

• Three workflows of increasing complexity that share one audit backend
• Claude as a proper agent with tool-use, not as a single-shot prompt
• Postgres-backed run log with structured error paths
• Single Docker Compose stack that deploys to any VPS

The solution

• Workflow 1 (simple): GitHub Activity Digest every morning across configured repos, posted to Slack
• Workflow 2 (moderate): Content Distribution Agent drafts Twitter thread, LinkedIn post, and newsletter blurb from each new blog post, queued for review
• Workflow 3 (advanced): Gmail Inbox Ops Agent routes incoming support mail through Claude with four bound tools (customer lookup, KB search, draft reply, file bug in Linear)
• Shared workflow_runs table so a single dashboard covers all three
• n8n on port 5678, Postgres on 5451 (host) with pgcrypto enabled

My role

• → Workflow design and audit schema
• → Claude tool-use wiring for the inbox agent
• → Docker Compose stack with port allocation that avoids conflicts with other local stacks
• → Dashboard over workflow_runs for unified observability

UI direction

n8n's native editor for workflow design, Postgres for state. Dashboard is deliberately minimal: runs in, status, timings, errors. No dashboard frills.

User flows

Key learnings

• Postgres audit beats opaque platform logs once you debug a real production flow once
• Claude agents with explicit tool schemas are more reliable than prompt-chained pipelines
• A single shared audit table across workflows is the single best design decision for an agent studio
• Production n8n is a different craft from the tutorial-grade workflows that dominate Fiverr listings