sloki9637 · one-person studio
sloki9637 is a one-person studio: planning, code and operations all come from the same person. I build the things I wanted to use myself, like a mood diary, an RSS reader and a saju app, ship them to the stores, and keep fixing them afterwards. Developer tools I needed along the way go out as open source under the Theorvane name.
Apps
Built with Flutter, Expo and Electron. Screen design, app code, the NestJS servers and store review are all handled by one person. Open a card for screenshots, a longer description and store links.
Pick today's mood and a tile lands on a 2D path. Keep going and a little dog walks the road you've built.
Builds your saju chart from solar terms and true solar time, and lets tarot cards answer the story you type. No login.
A video editor where your footage never leaves your machine. Tell the agent next to the timeline what to do and it cuts, trims and exports.
Add RSS feeds, get new posts in one place with AI summaries, and see at a glance what to skim and what to actually read.
Keep a record of films, books and shows, and turn a passing idea into a finished piece of writing with AI.
Read GitLab merge requests, diffs and pipelines on your phone and approve from there. Self-hosted GitLab works too.
A 4v4 snowball capture-the-flag the whole family can play. Freeze the other team with water blasters and bring their flag home.
Open source
When something was missing while building the apps, I built it and put it on GitHub: a library that declares MCP servers with decorators, a GitHub Action that reviews Terraform plans, and so on. OpenScene and LabFox ship under this name too. Each README says plainly what works and what doesn't yet.
TypeScript library
Declare MCP servers with TypeScript classes and decorators. Definition validation, MCP SDK compilation, stdio and Streamable HTTP, and a LangChain adapter are included; hosting and authorization stay with the application.
TypeScript library
Write LangChain JS tools and agents with @Tool() and @Agent() decorators. LangChain does the running; TypeChain only keeps types and schemas honest. There's also a bridge that turns a TypeMCP server into tools.
GitHub Actions
Review Terraform plans inside GitHub Actions. Risk and pass/fail are decided by code; the AI only explains a plan with sensitive values removed. No hosted service or GitHub App required.
MCP server
A local stdio MCP server for the Proxmox VE API. Read operations and state-changing operations are separated, and risky actions run only when explicitly allowed. Shipped as a checksummed GitHub Release archive.
How
No grand principles. Just three habits that stuck from working alone.
Screen design, app code, API servers, CI pipelines, store review and support replies are all one person's job. Decisions are quick, but there are only two hands, so I cut features and keep things small.
I don't wait until it feels done. The core goes out first, and usage logs and store reviews decide what the next version fixes. When a review gets rejected, the reason gets written down too.
Saju AI has no login and tells users apart by a single device key, LabFox's tokens never leave the device's secure storage, and OpenScene uploads footage nowhere. Data I don't collect is data I don't have to protect.
Lab
Small things I run before they reach a store. Right now that's Velra, a random chat a handful of people are trying with me.
Blog
What went wrong at launch, why a review got rejected, tools that turned out to be worth it. It all ends up on the blog.
Bug reports, feature ideas and "want to build something together" all go through the same form. I reply as soon as I've read it.
