April 4, 2026 · 8 min read

I Productized My AI Memory Server and Sold It for $29

I've been running a personal semantic knowledge base for over a month. 105+ thoughts captured. It remembers everything I tell it across every AI tool I use. On a Friday night, I decided to package it up and sell it. Here's the whole story.

The Problem That Started Everything

I use AI tools constantly. Claude Code for building. ChatGPT for brainstorming. Cursor for quick edits. And every single one of them has amnesia.

I'd have a great conversation with Claude about an architecture decision, switch to ChatGPT to think through a different angle, and... nothing carried over. Every tool, every session — a blank slate. I was repeating myself constantly. Explaining the same project context. Re-describing the same preferences. It felt like working with a brilliant colleague who gets a full memory wipe every morning.

Then I found Nate Jones's Open Brain — a video walkthrough for building a personal AI memory server using Supabase, pgvector, and MCP. The idea was simple: one place to store "thoughts" with semantic search, accessible from any AI tool via MCP.

I built it. And it changed how I work.

What Open Brain Actually Feels Like

Here's what my setup does: I tell Claude something important — a decision, an insight, a person's name and what they're working on — and it gets captured as a "thought." Automatically tagged with topics, people, action items, dates, and type. Embedded as a 1536-dimensional vector for semantic search.

Later, I'm in a completely different tool, different context, different day. I ask "what did I decide about the database?" and it finds the right thought — even if I never used the word "database" in the original capture. That's semantic search. It searches by meaning, not keywords.

"Your AI tools forget everything between sessions. Fix that in 10 minutes."

— the tagline I ended up with, because it's literally the pitch I make to myself every day

After a month of daily use, I had 105+ thoughts stored. Architecture decisions, project context, people notes, ideas for future projects. All searchable from Claude Code, Claude Desktop, ChatGPT, Cursor — anything that speaks MCP.

The "Why Don't I Sell This?" Moment

I've been wanting to build passive income for a while. No freelancing, no clients, no ongoing people interaction. Just build once, sell forever. A Friday night mood.

I looked at what existed: ContextForge, Memora, AMP, Hoofy, Phloem, Engram, Memobase, Pieces MCP — tons of open-source GitHub repos on Hacker News. Every few weeks someone posts "Ask HN: MCP server for long-term memory?" and gets a pile of links to scattered repos with no docs.

The pattern was obvious: nobody was selling this. Everything was free but broken. A YouTube video here, a README there, a GitHub repo with 3 stars and last updated two months ago. People would follow Nate's guide, hit friction on deploy, auth, or capture channels, and give up.

Meanwhile, in Nate's community, I could see people building real things on top of Open Brain:

• One person built a full vertical SaaS for business brokers on top of it. 135 thoughts in, called it "a real competitive moat."
• Someone else hooked it into Home Assistant. AI memory for your smart home.
• People asking about OAuth, token burn control, how to get started — real demand, real friction.
• Non-programmers building with Claude Code. The audience isn't just devs.

The gap wasn't "does this need to exist?" It was "does a polished, documented, just-works version need to exist?"

Yes. Obviously yes.

The Genericization: 5 Edits

Here's the part that surprised me: my code was already 95% generic. Open Brain's architecture is clean — it's a Deno + Hono server with MCP SDK, calling OpenRouter for embeddings and metadata extraction, storing everything in Supabase Postgres with pgvector.

Total changes to make it a product:

1. Server name: "open-brain" → "brain-kit" (plus an optional BRAIN_NAME env var so people can call it whatever)
2. Tool description strings: "Open Brain" → "your Brain Kit"
3. Default REST source: "slack" → "api"
4. Supabase config: project ID and function path renames
5. Folder rename: open-brain-mcp/ → brain-mcp/

That's it. The SQL migration, all MCP tool logic, the embedding pipeline — shipped as-is. ~400 lines of TypeScript. The code wasn't the hard part.

The Hard Part: Everything Around the Code

The code took 5 minutes to genericize. The product took a night to build. Here's what actually ate the time:

Documentation

A README that takes someone from zero to deployed in 10 minutes. Prerequisites, env var table, 5-step quickstart, architecture diagram, troubleshooting, FAQ. Plus 4 separate AI client connection guides (Claude Code, Claude Desktop, ChatGPT, Cursor) and a webhook API reference.

This is where most open-source MCP projects fail. The code is fine. The docs are "clone this repo and figure it out." I wanted someone to download a zip, run ./setup.sh, and be capturing thoughts in 10 minutes.

Capture Channel Recipes

Five ways to get thoughts into the system:

• MCP — the core. Works with Claude Code, Claude Desktop, ChatGPT, Cursor, anything MCP-compatible
• REST/Webhook API — generic HTTP endpoint, curl it from anywhere
• Slack + Zapier — instructions, no code needed
• Browser bookmarklet — a JS one-liner that captures the current page title and URL
• iOS Shortcut — capture thoughts from your phone's share sheet

Each one has a guide. Not "here's an API endpoint, good luck" — actual step-by-step with expected outputs.

The Deploy Script

One command: ./setup.sh. It links to your Supabase project, pushes the database migration, and deploys the Edge Function. If your .env is filled in, it just works. I tested this from a fresh Supabase project and it deployed in under 2 minutes.

The Friday Night Blitz

I won't pretend this was a carefully planned, weeks-long project. Here's the actual timeline:

From "not sure about passive income" to a branded product on a custom domain with AI-generated art and a Gumroad store. One Friday night.

The Stack

For the curious:

• Brain Kit server: Deno + Hono + MCP SDK (TypeScript)
• Embeddings: OpenRouter (text-embedding-3-small, 1536 dimensions)
• Metadata extraction: OpenRouter (LLM extracts topics, people, action items, dates, type)
• Database: Supabase Postgres + pgvector (HNSW indexes)
• Deploy: Supabase Edge Functions (Deno runtime)
• Landing page: Static HTML on Vercel. No framework. No build step. Just HTML, CSS, and vibes.
• Icons: Generated with Google Imagen 4 API (watercolor style)
• Cover art: HTML templates → Puppeteer screenshots (radial gradient masks to blend the watercolor brain into dark backgrounds)
• Domain: dashbuilds.dev on Vercel ($30/yr)
• Analytics: Vercel Web Analytics (free tier)
• Payment: Gumroad (no monthly fee, just transaction %)

Running cost for buyers: ~$0.30/month on Supabase free tier + OpenRouter usage. That's it.

What I Learned

1. The code is the easy part

Genericizing Open Brain was 5 edits. The value isn't in the TypeScript — it's in the documentation, the deploy script, the capture recipes, the client guides. It's the difference between "here's a GitHub repo" and "here's a product that works in 10 minutes."

2. The market validates before you ship

I didn't need a landing page to know this would sell. People in Nate's community were already building businesses on Open Brain. People on Hacker News were posting "Ask HN: MCP server for long-term memory?" every few weeks. The demand was loud. Nobody had a polished answer.

3. $29 is the right price for zero reviews

No testimonials, no social proof, no track record. $29 is low enough that someone thinks "even if it's mediocre, I'll save more than $29 in time not figuring this out myself." The plan: $39 after 20 sales, $49 after 50.

4. AI-generated assets are good enough

The watercolor brain icon, the geometric lines, the cover art — all AI-generated. Google Imagen 4 for the icons, HTML + Puppeteer for the Gumroad images. I spent zero dollars on design. The result looks like it was made by a design agency. (Okay, a design agency that likes purple a lot.)

5. Ship the smallest thing that's useful

Brain Kit doesn't have a web dashboard. No multi-user support. No email capture channel. No Telegram bot. No custom embedding models. The v1 is: server + migration + deploy script + docs + 5 capture recipes. That's the product. Everything else is v2.

What's Next

I'm building dashbuilds.dev into a home for all the tools I make. Brain Kit is the first one. More coming — all the same philosophy: deploy-ready, documented, designed to just work.

For Brain Kit specifically:

• OAuth guide (the community is begging for this)
• Email capture channel (Resend inbound parsing)
• Telegram bot recipe
• Multi-user support with Supabase Auth + RLS
• A demo GIF showing it in action (should 2-3x conversions)

If you're in the MCP/AI tools space and want to try it, Brain Kit is $29 on dashbuilds.dev. You get the full source, MIT license, lifetime updates. Deploy in 10 minutes. Your AI tools will finally remember.