Self-host licence · v1
The system that took five products from a Reddit thread to a live Stripe storefront with scheduled video. You license it once and host it yourself.
Bring your own agent. Every step is a shell command over JSON state on disk, so Claude Code, Codex, Cursor or a plain terminal can drive it. The 11 subagents and 13 skills in the kit are written in Claude Code's format. What that does and does not mean.
One-time payment, no subscription. It runs on macOS and needs your own Claude Code subscription plus Cloudflare and Stripe accounts before it does anything. Every one of those bills is in the table below.
$ ap step product-06 3 done✗ blocked — site/index.html has 4 unfilled placeholders✗ blocked — offer.amountCents is empty $ ap step product-06 3 done✓ step 3 done · 2 outputs written · run logged
§ 01 bring your own agent
Every step is a shell command that reads and writes JSON files on disk. There is no GUI and no proprietary runtime. If a thing can run ap next product-06 in a terminal, it can run this: Claude Code, Codex, Cursor, Windsurf, an API agent with a bash tool, or you.
ap is zero-dependency Node. ap next prints the current step and what is still owed, ap step … done runs the gate, and all of it lives in project.json and the scratch folder. An agent can ask the CLI what to do next, and a run that stopped halfway picks up where it left off.
They are written in Claude Code's format, and Claude Code is the path this was built and run on. Other agents do not load .claude/skills/ or .claude/agents/ by themselves. The files are plain markdown, so you can hand them to another agent as instructions or port them, but that part is your work.
.mcp.json registers Cloudflare, Stripe, Replicate and Higgsfield over OAuth. MCP is an open standard, so any MCP-capable client can connect to those same servers. That includes ChatGPT and other agents with MCP support, not only Claude.
Claude Code is the tested path and the only one that has done a launch end to end here.
Another agent will drive the same commands and read the same state, which is what the design is for. The configuration is not free, though. You point it at the skills yourself, and nobody has yet run all six steps that way, so budget time for setup.
§ 02 what a launch costs when you assemble it
Launching one small digital product properly means market research, a landing page with working payments, the product itself, short-form video, and scheduled posting. Every one of those has decent tooling. Nothing joins them up, so you do it yourself, by hand, between five invoices.
That is the realistic midpoint for one launch bought from freelancers and tools. Keeping one product's marketing running afterwards costs another $450–$1,550 a month.
Line items and sources in the table below.Across 146,271 Gumroad products, 44% earn nothing at all. A $4,000 launch cost has to clear before anything else does, and most launches never get there.
Source: State of Gumroad 2026, insightraider.com.The page says four scripts; the zip has twelve. The video hook does not match the headline, and the domain in the meta tags was never bought. No tool catches any of it, because the check spans two tools.
Every one of those shipped here before the gates existed.The closest thing to step one that ever existed as a product was GummySearch. It shut down on 30 November 2025 after Reddit revoked commercial Data API licensing.
That is why the research step here drives a real Chrome session instead of an API. It is also the most fragile part of the whole system. Read the FAQ before you buy. Source.
Every figure above is a published market rate with a link, not a projection. None of it says what you will earn. The full working, including the figures that could not be sourced, is in the research this page was built from.
§ 03 total cost of ownership · read this before buying
This is the strongest argument against buying, so here it is in full. The licence is a one-time fee for the pipeline, the agents and the gates. Every vendor below still bills you separately, and none of those accounts are included, resold or discounted here.
If you do not already have a Mac, a Claude Code subscription and your own Cloudflare and Stripe accounts, this is not for you, and no discount fixes that.
What you are buying is the state machine that joins the six steps together and refuses to advance when the work is not actually done.
§ 04 how it works
Download the kit, unzip, run the setup guide. The CLI has zero runtime dependencies; npm ci is only there for the test runner. Then ap new <slug> scaffolds a project folder with all six step directories.
ap next <slug> tells you the current step, which skill runs it and what a human still owes. You open a Claude Code session, run that skill, and it delegates the long work to the role-scoped agents. State lives in project.json, not in the chat, so tomorrow's session starts where today's stopped.
ap step <slug> N done checks the work before it records it. The step's outputs must exist on disk, contain no unfilled placeholders, and the fields the next step reads must be set. If any of that fails, the CLI prints the exact reason and leaves the step open.
§ 04b the six steps
◆ a gate sits between every pair of these
Steps 1 and 4 need you at the keyboard. Reddit access goes through a real Chrome session and a human clicks the captcha; the three social signups are done by hand on purpose, because automating a platform's signup form is how the account gets banned six months later, once it matters.
§ 05 the offer
Self-host licence
$1,499
one time · instant gated download · 86 files, ~208 KB
ap CLI, which holds state, the six steps, the gates and the run ledger, with zero runtime dependencies.mcp.json registering four remote MCP servers (Cloudflare, Stripe, Replicate, Higgsfield) over OAuth, with no keys in the file30-day refund, no questions asked. Email the address on your receipt and say the word refund. You keep the files; I am not going to argue about a zip.
No revenue claim is made anywhere on this page. The kit ships a product; whether anyone buys that product is down to you. Three of the five products built with it are still commercially unproven.
Done-for-you setup · enquiry only
$3,500
not open yet · no payment taken here
Your address is stored for this one thing and nothing else. No list, no sequence.
§ 06 the awkward questions
ap CLI (state, steps, gates, run ledger), 11 role-scoped agents, 13 skills, the project templates including the Stripe Pages Functions and the paywalled download, nine scripts, four documents led by a 592-line architecture write-up, 30 passing tests, a single-operator commercial licence and a setup guide written for this kit. No accounts, no credits, no support, no hosted version.Mostly yes, with one caveat. The pipeline is a command-line tool over JSON files: ap is zero-dependency Node, every step reads and writes state on disk, and nothing is locked behind a GUI or a proprietary runtime. Any agent that can run a shell command can drive it, so Codex, Cursor, Windsurf, an OpenAI or Anthropic API agent with a bash tool, or you in a terminal all work. Because the state is inspectable, whatever drives it can ask the CLI what to do next, and a half-finished project is picked up where it stopped.
The caveat is the .claude/ folder. The 11 subagents and 13 skills are written in Claude Code's format, and only Claude Code loads them by itself. They are plain markdown, so another agent can be pointed at them as instructions, or you can port them to whatever format it uses. That is work you do, not something the kit does for you. The .mcp.json is different: MCP is an open standard, so any MCP-capable client, ChatGPT included, can connect to the same four servers.
Claude Code is the tested path and the one all five products were run on. Nobody has taken another agent through all six steps, so the design supports it and nothing yet proves it.
LICENCE.md inside the zip. If that matters to you, email me before you buy and I will send you the licence text first.