The automation that refuses to fail quietly.
n8n already runs workflows. The itch Forge scratches is everything around that: deciding what to build from a sentence, proving it's safe before it ships, and — the part I actually cared about — noticing when a run breaks and repairing it without me.
It's a platform for people who'd rather supervise a system than babysit a pile of cron jobs.
A build in seven beats
January 2026
The 3am flameout
A scraping automation died silently for a week and I only noticed when the data went stale. The itch: what if the automation could notice — and fix itself?
February 2026
Scoping & the four layers
Wrote the first ADRs. Picked n8n as a swappable engine rather than building an executor from scratch. Drew the boundary: dashboard, stateless agent layer, data, engine.
March 2026
Foundation + the LLM gateway
Docker Compose stack, Postgres schema with Alembic, and the single gateway every model call flows through — cost logging, retries, and a hard daily budget from day one.
April 2026
The generator — and the week validation ate me alive
Turning English into workflow JSON was quick. Trusting that JSON was not. Half-valid graphs, phantom node types, a delete step with no approval flag. The validator became the real project.
May 2026
Monitor + diagnostician
Run ingestion, health computation, and the agent that reads a failure, writes a root cause, and proposes a patch — gated by a risk rubric so risky fixes wait for a human.
June 2026
Dashboard + hardening
The Next.js control surface, then the unglamorous parts: rate limits, backups, structured logs, and a load test to prove the stateless design actually scales.
July 2026
Polish & launch
Docs, a demo mode that runs with zero backend, and the site you're reading now.
The problems worth solving twice.
Never trust generated JSON
The validator statically checks node types against the engine, resolves every connection, verifies credential references, and blocks destructive actions unless an explicit approval flag is set. Generation without validation is a liability, not a feature.
Repair without recklessness
Auto-repair is gated by a risk rubric (ADR-004). Low-risk patches apply automatically; anything that touches credentials, deletes data, or rewrites logic waits for one-click human approval. Every action hits the audit trail first.
Generation can't block the API
Workflow generation is an LLM round-trip — too slow to hold a request open. It runs as a Redis-queued background job, so the API stays snappy and the work survives a restart.
Nothing is ever silently destroyed
Overwriting or deleting a workflow always writes a versioned backup and an audit entry first. Rollback to any prior version is a single API call.
Why each piece is here
Mature execution engine — kept swappable behind an adapter so the brain never depends on it.
Typed, async, and self-documenting; schemas double as validation.
Durable state and audit in Postgres; the job queue and cache in Redis.
One codebase for the marketing site and the dashboard, deployable to Vercel.
A capable default with a local fallback — and one gateway costing every call.
The whole platform comes up with a single command.
Want to build something — or collaborate?
The source is on GitHub. If you're working on agentic systems, workflow tooling, or just want to trade notes, I'm easy to reach.