Skip runtime LLM calls: let Claude generate your no-code automations via MCP instead
Potential-Art7696 · reddit · 2026-09-30
A solo developer shares UluP Spaces' architecture decision: the no-code automation editor (triggers + declarative JSON actions) almost got an askAI-style node that would call a model on every real-world event — costly forever and requiring async execution.
Instead, they flipped the problem: extended the product's existing MCP server with list/test/create automation tools, so users describe the automation to Claude in chat, Claude dry-runs it (no real data touched), and creates it. Created automations start disabled and reuse the same validation and execution engine as the visual editor — only who writes the JSON changes. Model costs fall on whoever is chatting, not the infrastructure per event.
A concrete example of trading runtime AI for config-time AI, with the author soliciting alternative takes on the trade-off.
More from coding & agent
- Agent Now Picks Its Own Effort and Hands Off to Subagents, Leading Cost-Performance on DeepSWE and Terminal-Bench 4.0 — _jzhao · 2026-09-30
- Hamel Husain: Four Ways to Run AI Evals When Traces Contain Sensitive Data — HamelHusain · 2026-09-30
- OpenAI DevDay live thread: 21 updates, headline release likely the always-on agent — kimmonismus · 2026-09-30
- Apodex 1.1 launches: agents can re-plan mid-task without restarting research — SucceededMind · 2026-09-30
- Deterministic memory layer for compliance agents: rules before LLMs, AI never changes verdicts — Friendly-Hat9209 · 2026-09-30
- Neuroscience Researcher Uses Claude and GPT as Rival Research Assistants — Now Unsure Where to Publish — neuroecology · 2026-09-30