AI Systems

Articles and newsletters on AI Systems: agent architecture, n8n, MCP, orchestration, self-hosting, monitoring, security, human control points.

Nothing published under this topic yet.

Catalogue

Products

AI Systems
MCP Server With OAuth 2.1
Your MCP server works when you try it once. Here is what it does in real use. One shared key in every client configuration cannot be revoked, cannot be attributed and never expires. What you get A credential ledger you can switch off, one client at a time OAuth 2.1 with PKCE and refresh rotation, built on RFC 7636 and RFC 7009 Self registering clients that still need a human to approve them Tokens bound to the one server they were issued for A stateless Streamable HTTP transport and a tool structure that scales The runbook for the day a token outlives its client How it is written Every chapter opens with how it breaks and ends with the one decision that stays with a person. Who it is for Builders who run an MCP server in real use, or are about to. Format A PDF e-book of 71 pages, 6 by 9 inches. Licensed for personal use. The file is protected: copying, editing and printing are restricted. Stellarwind AI publishes how real AI systems are built, set up and run.
Value$97
AI Systems
MCP Tool Design And One Registry
A tool an LLM cannot read is a tool it will misuse. The model sees only what you wrote: a name, a description and a schema. Everything else it guesses. What you get Three rules that make a tool description something a model can act on A ban list for vague verbs, and a policy of no delete tools One registry that declares each operation once and derives REST, OpenAPI, MCP and CLI Parity tests that prove the four never disagree The runbook for the day a tool is called wrongly How it is written Every chapter opens with how it breaks and ends with the one decision that stays with a person. Who it is for Builders who expose tools to a model and want fewer wrong calls. Format A PDF e-book of 49 pages, 6 by 9 inches. Licensed for personal use. The file is protected: copying, editing and printing are restricted. Stellarwind AI publishes how real AI systems are built, set up and run.
Value$47
AI Systems
Agent Orchestration With Deterministic Guardrails
Let the model interpret. Never let it decide alone. An agent system fails in four ordinary ways: a run resumes twice, an approval nobody sees expires, a model call is submitted again after a timeout, and an agent finds a second door around the check. What you get A typed split between what the model interprets and what code decides Leases and atomic checkpoints for runs that can crash Approvals bound to the exact thing that was reviewed A gate that turns intent into action, with no second door Receipts for every provider call, and replay from what the provider actually said A safe way to drive local AI command line tools as engines How it is written Every chapter opens with how it breaks and ends with the one decision that stays with a person. Who it is for Builders who run agents that act on real systems. Format A PDF e-book of 66 pages, 6 by 9 inches. Licensed for personal use. The file is protected: copying, editing and printing are restricted. Stellarwind AI publishes how real AI systems are built, set up and run.
Value$147
AI Systems
Multi Tenant Postgres, RLS And Safe Migrations
One missing clause in a policy is one customer reading another customer’s data. A policy can pass every test and still leak in real use. What you get A tenant id the database sets for you, not the application Row level security policies you have attacked from a second tenant Privileges granted in the order that cannot leak Migrations that widen before they tighten Rehearsals inside a transaction that rolls back A database move you can prove with a schema hash and row counts How it is written Every chapter opens with how it breaks and ends with the one decision that stays with a person. Who it is for Builders who run a multi tenant product on Postgres. Format A PDF e-book of 65 pages, 6 by 9 inches. Licensed for personal use. The file is protected: copying, editing and printing are restricted. Stellarwind AI publishes how real AI systems are built, set up and run.
Value$77
AI Systems
Cloudflare As Application Infrastructure
One Worker, one job, one host. A setup works until the origin is reachable around the proxy, or an email link changes state because a prefetcher opened it. What you get One Worker per job, each with its own host and its own bindings A D1 outbox that your backend drains and applies exactly once R2 buckets you can lock and restore from Links in email that never change state on GET An origin that only Cloudflare can reach Rate limits keyed on the client you can trust How it is written Every chapter opens with how it breaks and ends with the one decision that stays with a person. Who it is for Builders who run an application on Cloudflare Workers, D1 and R2. Format A PDF e-book of 60 pages, 6 by 9 inches. Licensed for personal use. The file is protected: copying, editing and printing are restricted. Stellarwind AI publishes how real AI systems are built, set up and run.
Value$77
AI Systems
Self Hosting AI On Hetzner With Coolify
Your stack, your server, your rules. Including the ones that page you. A server you host yourself is yours to keep alive: the backup nobody restored, the container that eats the host. What you get A Hetzner server reached through Tailscale Coolify with a migrator that keeps the owner credential away from the service A Cloudflare only origin you can prove Backups to R2 that you prove by restoring them A monitor that pages you on Telegram Limits that stop one container from eating the host, and the incident runbook How it is written Every chapter opens with how it breaks and ends with the one decision that stays with a person. Who it is for Builders who host their own AI stack and carry the pager. Format A PDF e-book of 58 pages, 6 by 9 inches. Licensed for personal use. The file is protected: copying, editing and printing are restricted. Stellarwind AI publishes how real AI systems are built, set up and run.
Value$97
AI Systems
The AI Operating System
Stop assembling it from ten blog posts. Six books make six layers correct on their own. What remains breaks where two layers touch: an audit that finds nothing, a restore that undoes a revocation. What you get The map: how identity, tools, orchestration, data, the edge and the servers fit together The interfaces between the layers, and what fails when the two sides disagree The order to build the layers in Seams you can test, each with the check that proves it One owner and one human decision for each layer A one page runbook that works across layers How it is written Every chapter opens with how it breaks and ends with the one decision that stays with a person. Who it is for Builders who run a live AI stack and own more than one layer of it. Format A PDF e-book of 72 pages, 6 by 9 inches. Licensed for personal use. The file is protected: copying, editing and printing are restricted. Stellarwind AI publishes how real AI systems are built, set up and run.
Value$197
AI Systems
Production AI Stack Map
See the whole stack on one page. The layer nobody owns shows up as an incident. This guide makes it show up as an empty cell first. What you get Eight layers of a production AI stack, from the edge to the runbook For each layer: what it is for, and where it runs, described as a role The failure to rule out first, and the decision that stays with a person A page you copy and fill for your own stack A pointer to the book that goes deep on each layer How it is written Every chapter opens with how it breaks and ends with the one decision that stays with a person. Who it is for Anyone who runs an AI stack and wants to see all of it on one page. Format A free PDF guide of 27 pages, 6 by 9 inches. Licensed for personal use. The file is protected: copying, editing and printing are restricted. Stellarwind AI publishes how real AI systems are built, set up and run.
Free