Recomby.ai

01 The Realization: Build a Plugin, Not an Agent

I spent a year on this project, and most of it was spent hitting walls. I wanted to build an AI agent that could do real professional work. First I tried modifying an existing one, NanoClaw — but it was unstable and its ecosystem was too small to bet on. So I stepped back and tried building one from scratch, which was worse: the foundational capabilities of an agent are being pushed forward by the big labs with an entire ecosystem behind them. I couldn't keep up alone, and there was no point — they'd already built it, for free.

After hitting the wall twice, it clicked: I shouldn't be building an agent. I should be building a plugin.

What makes a plugin the right shape is that it isn't locked to any one agent. The same plugin installs on Claude Code, on Codex, on anything else — I don't care what's underneath, because the ecosystem keeps making agents stronger, and my plugin empowers all of them at once. The more agents there are, and the better they get, the more my plugin is worth.

That frees up all my energy for the one thing that actually matters: understanding a single trade deeply enough to do it better and faster than anyone else. A capable agent is still a blank slate in any specialized field. I take a business I've truly absorbed, turn it into a plugin, and the moment an agent installs it, it becomes an expert in that trade. I don't build the agent — I make the agent know its craft.

And the first trade I picked was the agent economy itself. As AI becomes how people find information and complete transactions, every business runs into two new problems: how to be seen by AI, and once seen, how to actually close the sale. My two open-source plugins map onto exactly those two things.

02 Part One: GEO — Getting You Found by AI

GEO (Generative Engine Optimization) is about being seen. When a large model answers a question, it synthesizes a single best answer across many sources — and what you're competing for is to be cited in it.

It's a fundamentally different contest from SEO. SEO is a contest of who optimizes better — stacking keywords, accumulating backlinks, shoving your link toward the top of the results so the user sees a list and clicks through themselves. GEO is a contest of who has the real substance: AI gives no list of links, it folds many sources into a single spoken answer, and what you're competing for is to be the basis of that answer. And what AI decides to trust isn't keyword density — it's whether the content carries first-hand business insight. Padding it with filler isn't just useless; it gets penalized.

I built this into an open-source plugin, recomby-geo: it packages GEO expertise into skills an AI agent can call directly, paired with a 7-stage collaborative workflow from intake to re-audit. MIT-licensed, zero external dependencies, zero API key — and past 450+ GitHub stars. Install it, and a general-purpose agent gains the ability to do GEO professionally.

But the deeper I got, the more one problem nagged at me — GEO alone still only gets you traffic.

03 Why Part Two: Without the Sale, GEO Is Just Traffic

AI mentions you in its answer, cites you — and then what? The user still has to leave, find your site, and place the order themselves. The chain breaks right there. Between being mentioned and being bought from lies a whole stretch of road no one has paved.

The real shift is this: once agents become the entry point, users will have the agent close the deal directly — compare, order, pay, all in one motion, with no platform middleman in between. If your business can't plug into that transaction track, then the visibility GEO gives you is just old traffic from a new source. Limited value.

So I built a second plugin.

04 Part Two: UCP — Getting You Transacted by AI

UCP (Universal Commerce Protocol) is an open standard backed by Google, Shopify, and 20+ partners that lets AI agents discover businesses and transact with them. ucp-onboard keeps the same shape — a plugin for agents — except this one teaches the agent how to get a store onto the transaction track: give it a merchant URL, and it runs the full onboarding pipeline — audit → generate the business profile → map the product catalog → configure checkout → validate and go live — until the store is "alive" on UCP, transactable by agents directly.

I didn't stop at using someone else's protocol. UCP today only defines buying products (SKU, price, inventory) — but commerce isn't only products. It's also services: consulting, design, the labor of an AI, on-demand SaaS. I submitted Issue #303 to the UCP consortium, proposing a Services Vertical that makes services agent-transactable too, with a full lifecycle from booked to in-progress to delivered to verified to settled. I'm not just using this protocol — I'm trying to extend it.

None of this is hypothetical. I ran the full pipeline against real e-commerce sites — from a small one with no standard format at all, correctly flagged as non-compliant, to a mature Shopify store that passed every check on the first try. It identifies and handles them correctly. It tests, it runs, it goes live.

05 The Delivery Model: An AI Employee, Not a Replacement

Both plugins land at the client as the same form — an AI employee. But the word I want to stress is employee, not replacement.

I thought this through carefully, to the point of writing it into the code: during content production, if the insight slots the business expert is supposed to fill aren't filled, the system hard-refuses to auto-fill them with AI content. The human has to stay in the loop.

Real substance can only come from someone who knows the business. So an "AI employee" here can only be collaborative — the AI carries the framework and the execution, and hands the insight piece back to you. For the client, the value of this is concrete:

• Cheaper than a full-time hire — and it actually does the work. The plugin connects to your office software (Feishu, DingTalk, Notion), pulls real business data, and carries out real tasks, not just dropping a report on you.
• The insight and the final say stay with you. Your business knowledge never passes to a third party, and key results are rendered for your review before anything moves on.

This "stay on the client's actual ground, get the work done with human and agent side by side" delivery model is exactly the spirit of FDE (Forward-Deployed Engineer) — not handing over a deliverable from across a gap, but the engineer and the agent going in together. I wrote that principle straight into the product as a hard constraint.

06 Finally, About This Path

There was no blueprint to any of this. Every step was forced out by the one before it — GEO got questioned as nothing but traffic, so I went and built agentic commerce to close the loop; I couldn't get an agent working no matter what, so I pivoted to plugins; and all along I kept trying different business models, hitting walls and reworking them.

Where it stands today: recomby-geo has crossed 450 GitHub stars, it's serving its first two clients, the Services Vertical proposal is in front of the UCP consortium, and the pipeline has run against real e-commerce stores end to end. None of these are big numbers yet — but every one is an artifact someone else can click and check. That, plus what this year taught me about how AI actually lands in the real world, is the return on the path. For how the four-gate GEO framework was derived from first principles and welded into the plugin, see First Principles Is a Scalpel, Not a Hammer. For why ontology — not a bigger model — is the real bottleneck to landing agents in the enterprise, see The Cheaper Tokens Get, the More Ontology Is Worth.