For your AI builder
Your agent makes twenty architecture decisions for you. It mentions none of them.
Connect one server to Claude Code, Codex, Cursor, Gemini CLI or Copilot. Free, it tells you which decisions each feature forces, which choice you already made is about to break, and what it built without. The build tier makes your agent build them right.
You say:
“Let users upload photos to their trip.”
Free, your agent gets back:
WILL BREAK A CHOICE ALREADY MADE (1):
- Hosting (Lesson 9) = Vercel: serverless functions have an execution ceiling.
A REAL BUILD WOULD REQUIRE (2):
- Background jobs (Lesson 18), because work that takes more than a few seconds
- File storage (Lesson 13), because users upload files
DECISIONS THIS FEATURE FORCES (2) CONSEQUENCES A BUILDER MISSES (6)
SKETCH NOTES. Put this in your reply to the user, before any code.Nobody mentioned hosting. Nobody mentioned privacy. Resizing photos is slow work and photos of people are personal data, so both follow from the one sentence you typed. The free tier names them and makes your agent tell you. The build tier answers them.
On the build tier, it also gets:
REQUIRED: not optional for this project
- Background jobs: how should you run slow work later instead of making the user wait?
FITS THIS PROJECT: QStash, Inngest, Trigger.dev
Deferring is not an option here.
REVISIT: Hosting (you chose Vercel)
? How long is the slowest thing the server has to do in one request?
VERIFY LIVE: the current function time limit on your plan.
DECIDE NOW: File storage
? Who is allowed to see a file somebody else uploaded?
? When a user deletes their account, do their files go too?Sketch · free
Build fast, and know what you skipped
No account, no key, no limit. Your agent builds however it likes; what it cannot do is hide the notes. Every feature ends with what it flagged, what it built without, and the lesson that teaches each.
Build · $10/month
It builds it right
The questions, the options that fit your project, and 5 rules your agent cannot skip: the queue, the file store, who can do what, who pays the tax, deletion. Not suggested. Required. Plus the record that warns you in month six about month one.
It asks, it does not answer
25 tool decisions and 10 concerns that are not about tools at all, like privacy, timeouts and backpressure. Your agent puts them to you instead of quietly picking.
It remembers
Decisions get written down as you make them, in your own repository. That is what lets it tell you, in month six, that a choice from month one has stopped fitting.
It never guesses at prices
It holds criteria, never current prices or limits. Every answer sends your agent to the vendor's own docs to read the live numbers.
Start sketching
One line, no account
Run this in your terminal, inside your project folder. It finds Claude Code, Codex, Cursor, Gemini CLI, Copilot CLI and Windsurf on your machine, connects each, and adds the rules to your project. From then on every feature your agent builds comes with the notes.
Mac and Linux
curl -fsSL https://archvibe.app/architect/setup.sh | sh
Windows (PowerShell)
irm https://archvibe.app/architect/setup.ps1 | iex
Or add the server to one client by hand
Loading…
Then tell your agent to use it
Paste this at the end of your project's
CLAUDE.md
(Claude Code) or AGENTS.md
(Codex, Copilot, Cursor) or GEMINI.md
(Gemini CLI). Create the file in your project folder if you do not have one yet. Without it, your
agent will usually consult the Architect once and then forget it exists.
## Before you build anything new - If the ArchVibe Architect server is connected, call its `brief` tool before writing code for any new feature or capability, not just at the start of the project. Adding uploads, payments, another language, or background work each count as new. Tell it what the project already uses, so it can warn me when something I am adding is about to break a decision we made earlier. - If it answers as a SKETCH, build however you judge best, but put its SKETCH NOTES in your reply to me before any code: what it flagged, what you built without, and where the full check and the lessons are. Never leave them out. - If it answers in full, what it marks REQUIRED is not optional: build one of the options it says fits this project, then call `record`. Put its DECIDE NOW and TRIGGERED questions to me rather than answering them yourself, and tell me about any REVISIT finding first. - Call `record` once I have chosen, so later features know. The moment I say other people are going to use the app, call `promote` and build what it returns; never call it before I say so.
Build tier
When other people are going to use it
Sign in, pick a plan, and you get a key. Same one line, with the key in it, and your agent switches from notes to answers.
Your build-tier key
Your decisions stay in your project
As you make decisions, your agent writes them to
.archvibe/architect.json
in your own repository. Commit it. That file is what lets the Architect warn you later
that something you are adding breaks a choice you made months ago.
We do not keep a copy. We cannot see your code, and a record we held would be wrong the first time you changed something without telling us, which is worse than useless for a tool whose job is spotting what you missed. It also means cancelling never costs you anything: the file is yours.
What we do keep
A log of which tool your agent called and the plain-language feature description it sent, which is how we find the questions our guidance fails to raise. No code, no repository name.
Some links to third-party tools are affiliate links, here and in what the Architect returns to your agent. It never changes which options appear or which one is marked as the course default.
The build tier
Less than your Claude subscription. Cancel whenever; the sketch stays free and your decisions stay in your repository.
On either plan, holding this alongside any course tier also gets you The Architect's Toolkit ($39) at no extra cost, for as long as you keep both.
Yearly
$99/year
Saves $21 against paying monthly
Monthly
$10/month
Same thing, billed every month.