Open source · MIT · built to solve my own problems first

Tools I built
to get my work done.

None of these started as products. Each one started as friction in my own week: a design system I needed off a website, a video that had to fit under a size cap, a timer that does not shame you for pausing it. I built the smallest thing that removed the friction, kept using it, then put it on GitHub under MIT in case it removes yours too.

Source for all nine is on github.com/Sagelakshya.

9public repositories
MITevery one of them
4languages, whatever fit
Local-firstno accounts, no uploads

The tools

design · video · workflow · attention
Python · FlaskMIT

control-center

One page for the actions an AI agent prepared but should not run itself: force-pushes, uploads to a live server, anything hard to undo. The agent writes the queue and shows every command verbatim, with the reason and the honest risk. It never presses Run. That stays with a person, by design.

JavaScriptMIT

snapkit

Capture any website's design system to disk: tokens, components, motion, assets, plus a CDP bridge that attaches to a real logged-in Chrome for pages a scraper cannot reach. Deterministic, no model calls.

PythonMIT

dubsync

Dub a YouTube video into another language with the voice kept in sync with the picture. Runs entirely on your own machine: Whisper transcription, natural Hinglish, one-click Windows download.

PythonMIT

focus-companion

A timer and a capture box in one always-on-top window, designed around an ADHD profile. Counts forward, never down to a clock time. Plain text files you own.

JavaScriptMIT

apply-autofill

A Chrome extension that fills job-application forms from a local JSON profile. Zero network calls, your details never leave your machine, and it never clicks submit for you.

PowerShellMIT

video-compress

Compress any video down to a target file size, say under 4 MB, with the least quality loss it can manage. One script, ffmpeg underneath, nothing uploaded anywhere.

PowerShellMIT

folder-notes

A sticky note for a Windows folder. Open the folder and the note is already there. Plain _NOTE.md files living inside the folder itself, so there is no database to lose.

PythonMIT

whatsapp-links

Turn a WhatsApp self-chat export, the place a lot of us dump links, into a deduped and categorised library. Python standard library only, deterministic, no model calls.

HTML · CanvasMIT

speedster

A game rather than a utility: a first-person POV speedster runner with a hand-written pseudo-3D projection and real physics from subsonic to the edge of light. Playable in the browser.

What they have in common

the constraints I kept choosing
01
It runs on your machine

No account, no upload, no server holding your files. dubsync transcribes locally, video-compress never sends the video anywhere, and apply-autofill makes zero network calls with your details.

02
Deterministic where it can be

snapkit and whatsapp-links do their whole job with no model calls at all. A model is a good collaborator and a bad dependency, so the mechanical work is code and the judgement stays with you.

03
Plain files you own

folder-notes writes a real markdown file inside the folder. focus-companion writes plain text. If the tool disappears tomorrow, everything it made is still readable without it.

Built alongside client work. If one of these is close to what you need but not quite it, the licence lets you fork it, and I am happy to talk through the design decisions.