Seven beans · One library · No SaaS
A small-batch shop selling written instructions that turn AI coding agents into something resembling employees. Markdown brewed at 200°F, poured over your repo. No subscription. No "platform." No condescension.
"Money never sleeps. The library compounds. Neither does the espresso machine."
House motto · Unofficial · Unapologetic
Your agent forgets your name between sessions and you call that "stateless architecture." This pack is the decision layer for what an agent remembers, forgets, and politely declines to mention. Episodic, semantic, procedural — and the policy discipline above all three. Stop letting your bot drown in its own state.
What's inside: 11 skills, 4 worked integration recipes (Redis, pgvector, Pinecone, multi-agent), 22 files, roughly 32,700 words. Install guide for Claude Code, Cursor, and Codex. Every file listed in the pack's MANIFEST.
Make N AI agents do your job at the same time. They work in separate git worktrees so they don't punch each other. They merge to main when local CI passes. You stay liable for the rebase. This is the actual pattern behind a multi-tenant SaaS built by one human operator who, statistically, was sleeping.
What's inside: 13 skills, 9 worked artifacts from a real setup, 29 files, roughly 34,000 words. Install guide for Claude Code, Cursor, and Codex. Skill 1 is free to read right now, in full, before you spend anything.
Two agents in a small room. One proposes. The other grades. They argue about prompts, RAG retrieval, evals, and copy until the numbers go up. You sleep. The bill goes up. The numbers go up faster. Convergence is the goal. Stopping conditions are included so you don't wake up to a $4,000 OpenAI receipt and a slightly better headline.
What's inside: 10 skills, 4 worked examples (prompt optimization, RAG tuning, eval-driven testing, strategy search), 21 files, roughly 16,000 words. Install guide for Claude Code, Cursor, and Codex.
Everyone is building a "GPT for X." You can do better. Turn your product's API into a production-grade MCP server — opportunity validation, tool schemas, auth, permission boundaries, prompt-injection threat model, Inspector testing, deploy paths, marketplace listing, monetization. Twelve skills. One server. No hand-waving.
What's inside: 12 skills, 3 worked server build-outs, 22 files, roughly 22,300 words. Covers stdio and Streamable HTTP transports, OAuth 2.1, and the Inspector test flow. Install guide for Claude Code, Cursor, and Codex.
You need 80 on-brand images. Your design agency quoted $22,000. Midjourney quoted "good luck staying consistent across 80 prompts, buddy." This is a Python CLI. Drop in a JSON of prompts. Get back a metadata-clean, web-optimized WebP library in fifteen minutes, for whatever your image model charges per scene and not a cent more. Tide in. Tide out. Idempotent.
What's inside: a 375-line single-file Python CLI, 8 skills, 2 complete example pipelines (22 finished scene prompts), 20 files, roughly 12,700 words. Post-processing runs on your CPU and costs nothing.
Read more →Already buying one of the four core packs? Add Memory at the regulars-only price. Fifty dollars off because you're not a tourist. (An agent without memory is a cause without consequences. We won't sell you that life.)
What's inside: identical to the standalone Memory pack. 11 skills, 4 worked integration recipes, 22 files, roughly 32,700 words.
All four core packs. Together. At a price your accountant will silently respect. Skip the bean-by-bean math: this is what the regulars buy on the first visit.
What's inside: 46 skills, 20 worked examples, 94 files, roughly 105,000 words across four zips.
The Flight of Four is the house pour. If you want less, or you want the whole shelf, there is a bag for that too.
Parallel Agent Lanes and the Auto-Research Loop, poured together. One runs your agents in parallel. The other makes two of them argue until the numbers go up. Buy the pair and the second cup costs less than the walk-in price.
What's inside: 23 skills, 13 worked examples, 50 files, roughly 50,000 words.
Everything on the menu that comes in a bag. All four core packs plus the Image Pipeline. The Flight of Four with the cold brew added. This is the shelf, not the sample.
What's inside: 54 skills, 24 worked examples, 114 files, roughly 117,700 words, plus the Python CLI.
Cupping is what a roaster does before they will vouch for a lot: sit down, grade it honestly, write up what is actually in the cup. Two audits, done by hand, single pass. You buy, you tell us where to look, and a report lands in your inbox within five business days.
How visible is your brand when someone asks an AI instead of a search bar? We run the prompts a buyer would actually type, grade where you show up and where a competitor takes your seat, and hand back the map.
Point us at your AI agent or MCP server and we cup it for the things that spill: prompt-injection surface, tool and permission boundaries, secret handling, the paths an attacker would actually walk. Then we write down how to close them.
The packs teach your agents a pattern. These are the machines themselves: source-only tools you drop straight into production and own outright. Heavier, built for the shop that runs its own line.
A deployable MCP server, not a lesson about one. Cloudflare Workers, Hono, Drizzle, and OAuth, wired and ready. The lockfile ships with it, so your first deploy is a clone away.
The Agent Memory pattern, cast as a library you can import today. A Node.js memory core, three storage adapters, and a migration utility. Drop it into an existing agent and give it something to remember.
Stop shipping RAG changes on faith. A Python harness on Ragas and DeepEval that scores your pipeline, catches the regression before your users do, and keeps the numbers honest between releases.
Yes. Deliberately. A skill is a plain text file your agent reads and executes, which means you can read the entire thing end to end before it ever touches your repo. There is no binary, no runtime, no daemon phoning home, nothing to audit that you cannot audit with cat. The value was never the file format. It is knowing which five prerequisites decide whether two agents can share a codebase, and what to do when the answer is no. We publish a complete skill from the Parallel Agent Lanes pack at /free, unedited and unlocked. Read it and decide for yourself.
Because a marketplace skill is one file and these are packs. Parallel Agent Lanes is 13 skills, 9 worked artifacts from a real setup, and 29 files. The MCP pack is 12 skills across transports, OAuth, permission boundaries, and a prompt-injection threat model. The comparison that matters is not one file against another file. It is the pack against one hour of a senior developer's time, which is what you will spend rediscovering the third prerequisite the hard way. Buy once, keep the files, v1.x updates are free for life.
More than most shops will let you. Every product lists its full manifest right on this page: exact skill count, worked-example count, file count, word count. One complete skill is published free at /free, the same file buyers get, nothing held back. Every pack ships a MANIFEST describing each file it contains. Everything is plain text. And if the zip still is not what you expected, reply to the delivery email with the word "refund" inside seven days and we send your money back without an interview.
A folder of markdown files plus worked examples. You drop them into your agent's context — Claude Code skills, Cursor rules, Codex AGENTS.md, any agent that loads a .md file. The pack teaches the pattern. The agent applies it. You own the files forever. Not a subscription. Not a "platform." A bag of beans you grind at home.
Because we already wrote the thing. You're going to pay us. The transaction is done. No monthly tax, no "your free tier is over" emails at 11pm, no surprise pricing changes when our investors demand growth. Buy once. Own forever. Sync to whatever machine you want. If we have to send you another invoice in six months, one of us is doing this wrong.
Seven-day refund, no interview, no exit survey. Reply to the delivery email with the word "refund" and we send your money back. After day seven we still listen, but the receipt is closed.
v1.x updates are free for life. New skill variant, bug fix, better example, fresh stack pattern — existing buyers get it. We'd only charge for v2.0 (significant restructure), and v1.x owners get a discount. We're not running a Subscription as a Service.
Yes. Single-user commercial license covers any commercial software you ship. The line we draw is on redistributing the files themselves or wrapping them into a competing skill product / paid course. Don't be that person. Build with them, ship something good, charge what it's worth.
There are exactly seven Kybalion hermetic principles and we have exactly seven packs. We thought that was funny. Hermes also happens to be the classical patron of merchants, messengers, and language — the literal job description of "a store that sells written instructions." We're not running a cult. We're running a coffee shop that read too much philosophy.
Because solo operators with AI agents are the new espresso machine, the new home roastery, the new "I do this myself, slowly, and it tastes better than the chain." That's the whole product. The metaphor and the offering rhyme.
You buy the audit like anything else on the menu. Right after checkout you land on a short intake form where you tell us where to look: your brand and competitors for the GEO Audit, or your repo or endpoint and scope for the Security Audit. We run a single pass by hand and send the report to your purchase email within five business days. One engagement, one report, no retainer.
A pack is a set of markdown instructions that teaches your agent a pattern; you keep the files and the agent does the work. A Roastery tool is the finished machine: source-only code you deploy or import directly, like the MCP server template, the Memory Kit library, or the RAG harness. Same shop, heavier gear. You own the source outright, single-user commercial license, same 7-day refund.