Brain in a Jar

I haven't crawled out from behind the computer in two months and my hands are shaking from the amount of coffee, but today I literally screamed across the whole house, because my agent, the one that lives inside my own app, went into the DaVinci Resolve knowledge base on its own, found the specific chapter about color correction in a manual that's 3500 pages long, and retold it to me in Russian, even though the manual is in English, and the only English I know is how to say fuck you, for fuck's sake.

How Is This Even

I'm not a programmer, just so we're clear, I've never hidden that, and every time I say it nobody believes me, because I've got a desktop app built on Rust and React that I assembled with the hands of Claude Code over two months, and today I bolted RAG onto it, a knowledge base system, from scratch, in one session, nine thousand plus lines of code, and it works, damn it, it actually fucking works.

Not "works in theory". Not "compiles and shows a blank page". I fed it the DaVinci Resolve PDF manual at 25 thousand chunks, an article from a website, subtitles from a YouTube video, and when I asked the agent "tell me what they're saying in the video", it called the search tool itself, formulated the query in English itself, picked the right source out of three itself, and gave me a structured answer in Russian. Without a single hint from me where to look.

The Team

I spun up four agents. Two coders, two reviewers. One coder was cranking out the Rust backend, the other the React frontend, each in their own git branch so they wouldn't step on each other's feet. The reviewers were checking the code and finding real bugs — I dunno, a SQL injection in LanceDB, an SSRF vulnerability in the scraper, a race condition in the temp directory for YouTube, type mismatches between front and back across five different data structures.

The craziest thing about it is how it iterates, because I'm sitting in my chair with a cup of coffee, saying "yeah that orange shit at the bottom of the sidebar, get rid of it", and three minutes later the coder commits, the reviewer checks, I see the result, and we move on to the next thing. In one day I went from "hmm let's think about which RAG framework to pick" to a fully working MCP server with four tools, settings, a dashboard card, PDF support via poppler, EPUB, web scraping and YouTube.

Model in a Binary

Inside the app there lives a tiny neural network 23 megabytes in size, which turns text into a bunch of numbers, a "vector", and texts similar in meaning produce similar vectors. It gets downloaded once on first use, after that it runs locally, no internet, no cloud, no subscriptions, right inside Rust via ONNX Runtime. When I realized this, I had a brief existential crisis, because three hours earlier I had no idea what embeddings even were, and now I've got a neural network built into my app and I'm discussing with the architect whether we should switch to the multilingual bge-m3 model for better search quality on Russian queries.

Killed Darlings

You know what I like most about this process — I started by drawing up a system with knowledge bases linked to chats, with an accordion in the sidebar, with automatic context injection, I wrote fifteen hundred lines of code, tested it, and deleted the whole fucking thing. Because once the agent had an MCP tool for search, all that manual linking became useless, it searches smarter on its own than any automation. Minus 341 lines in a single commit, and the app got better. I killed my darlings and I didn't regret it for one second.

Sadness

— personal, feel free to skip

I haven't seen anywhere on the internet anyone working with Claude Code like this, through a multi-agent team where an architect coordinates coders and reviewers, where agents hold their roles across the whole session, where you can launch even twenty parallel processes if the hardware can handle it. I see people using one chat, one session, "write me a function", and it's like watching a guy who bought a Formula 1 car and drives it to the bakery in first gear.

And nobody knows about it, because I don't care about marketing, I'd rather sit and tinker. Second month glued to the computer. Because it feels good.

The Numbers

For anyone curious, total for the session: 34 files, +9250 lines, 15 Rust modules (PDF/EPUB/web/YouTube parsers, a vector index on LanceDB, embeddings via fastembed, an MCP server with 4 tools, a settings system), 13 React components (knowledge base management, search, dashboard card, settings section). Two full audits with two rounds of fixes. One busted PDF parser, one SQL injection, one SSRF hole, all fixed before merge. Zero merge conflicts.

My agent can now search through my books, articles and videos, and answer me in plain human language. I'm not a programmer. I've got a knowledge base on directing that runs to several thousand pages. Tomorrow I'm feeding it in.