The cookbook

The cookbook is a set of markdown recipes your agent retrieves on demand. Recipes are written by us — humans, working from real deploys — and applied with your agent’s judgment.

What’s inside

One bundle, organized by topic: the standard deploy flow, a verify checklist, debug recipes, quota walls and how to negotiate them, day-2 operations, and multi-app / multi-service patterns. Stack-specific gotchas — a Next.js prerender quirk, an Astro preview-host allowlist, sizing a FastAPI + Celery pair — live inside those sections, so your agent doesn’t relearn them per deploy.

Your agent gets the cookbook in one nyx_brain call — the evergreen core plus the stack and topic recipes retrieval judges most relevant to your repo and briefing — alongside your project’s current state. There’s no follow-up retrieval step on the agent’s side; the matching happens server-side before the response.

Why static, not generated

We don’t ask an LLM to write the recipe at runtime. Three reasons:

• Predictability. A recipe that worked yesterday works today. No silent drift between deploys.
• Debuggability. When something goes wrong, we can read the exact text the agent saw and reason about why.
• Auditable, independently released. A recipe is fixed text, not a fresh generation, so we can pin exactly which version served a deploy — and update recipes on their own cadence, decoupled from the backend, with no redeploy and no agent rebuild on your side.

Judgment over orchestration

The recipe is a suggestion shaped by experience, not a fixed sequence. When your agent sees something the recipe didn’t anticipate — a non-standard build config, a missing secret, a registry the recipe doesn’t know — it deviates. We never force a state machine.

That’s the trade-off: outputs aren’t byte-identical across runs, but they’re outcome-equivalent. Two deploys of the same repo by the same agent both end with App Available, Ingress reachable, docs present.

How to read recipes

Recipes are agent-facing — written in the imperative second person, addressed to the agent reading them. You don’t browse them on this site.

If you want to see what your agent saw on a particular deploy, ask the agent — the cookbook is the nyx_brain tool result, and most clients let you inspect MCP tool results.

Next: Deploy a Next.js app →