n8n AI Workflow Generator

Describe in plain language what you want automated. Built an AI flow generator for n8n. Two minutes later a working automation drops out, ready to run, no engineer needed.

• 01User surface

  The whole experience the user sees

  This frame shows the live product: describe the workflow in plain language, get a ready-to-import n8n json · all nodes wired, credentials slotted, test data included. Every component is ours · scope, design, code, deploy.

• 02Stack behind the screen

  What's powering it: Python, TypeScript, Greenlet

  6 stack components run behind this frame · Python, TypeScript, Greenlet drive the visible UI; the rest sit in the data layer. All studio-owned.

• 03What we shipped

  Natural-language input · Hungarian and English both supported

  Workflow idea → running n8n flow in 2 minutes

• 04Status

  Private deploy · under NDA.

  Per the client's request the URL stays private · the build, architecture, and lessons can be shared in a scoping call.

• 01 · BRIEF

  Why does ChatGPT-generated n8n JSON never import?

  Diagnosed the failure mode: ChatGPT lacks node-schema awareness and credential semantics. We built a vector-search-first generator that always starts from a known-good template, then refines.

• 02 · ARCHITECTURE

  Stack decisions before any code.

  Decision doc captured the data flow, Python, TypeScript, Greenlet, FastEmbed role split, and the failure modes we'd handle in v1 vs defer. Cross-service boundaries (where AI ends and the web app begins) were drawn here so neither side leaked into the other later.

• 02 · BUILD

  Chainlit + FastEmbed + iterative refinement.

  Chainlit hosts the chat, the user describes the flow, FastEmbed retrieves the closest matching template, the LLM patches it to the user's spec. Iteration is conversational · 'add a Notion node after the Slack message' just works.

• 04 · POLISH

  Performance, accessibility, and observability.

  PSI / a11y / coverage budgets enforced as launch gates. Logging + metrics wired before cut-over · the team can answer 'is it working?' from a dashboard, not a Slack thread. Threat-model checklist signed off before traffic hits the box.

• 03 · SHIP

  Local-first deploy · history stays on the user's box.

  aiosqlite holds the conversation + template history locally · no external persistence, no leakage of the user's automation logic. Export button drops a JSON file ready for n8n's import.

• 01Chat

  Chainlit chat UI · iterate in conversation

  Same UI for first draft and refinement · users don't switch contexts to tweak.

• 02Search

  FastEmbed template index

  Vector search over a curated template gallery · the generator always starts from a known-good base.

• 03Export

  Import-ready n8n JSON

  Output is validated against the n8n schema before download · 'JSON copies cleanly into n8n' is a launch gate.

• 04Local

  aiosqlite-only persistence

  Conversation + template history stays on the user's machine · no SaaS lock-in for sensitive automation logic.

• 
  01Frame

  Init · 438 nodes indexed, quick-starts visible

  First message logs the boot tools (`Start n8n` / `Load node catalog` / `Build embeddings index` · 438 nodes). Quick-start chips (URL→Discord, RSS+AI digest, GitHub PR review, support triage) lower the barrier to first prompt.

• 
  02Frame

  Agentic build · UUIDs → JSON → validate → refine → import

  Agent loops through real tools: `generate 8 UUIDs` for node ids, builds the JSON, calls `validate workflow`, realises a Split Out node is missing, refines, validates again, saves, imports. Each step is a tool call you can drill into, not opaque LLM prose.

• 
  03Frame

  Handover · file location + import note

  After import, the agent shows the saved file path (`workflows/calendar_to_notion_via_openai.json`), notes about scheduling, format requirements, and the n8n sign-in caveat · the user has the artefact + the runbook in one place.

• 
  04Frame

  Second flow · 8-node order pipeline with branch logic

  More involved order-processing webhook · 8 nodes, VIP / non-VIP branches that have to converge into Postgres → SendGrid email → Asana task. Agent rebuilds when validation flags the connection logic, no human cleanup.