This is a submission for Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady
🌫️ Brume: A Little Mist Sprite That Breathes With You
You know that moment before an interview, an exam, or a difficult conversation when someone tells you to just relax?
I wanted to build something more helpful than that.
Meet Brume — a little mist sprite designed to make taking a pause feel natural. Instead of another productivity dashboard telling you to optimise your day, Brume invites you to slow down, breathe, and reset.
What I Built
Brume is a calming breathing companion with:
- 20+ breathing patterns to help you find a rhythm that works for you.
- Short guided sessions, including a three-minute “get ready” session and a one-minute alertness check.
- Optional camera feedback powered by MediaPipe, with pose processing on-device.
- An AI companion that can offer personalised insights using Llama 4 Scout 17B through Cerebras.
- Offline support, so the core breathing experience remains available after the first load, without requiring the AI.
The idea is simple: reduce the effort between feeling overwhelmed and taking a moment for yourself. No elaborate setup, no need to use AI, and no account required to start breathing.
Brume is still in development, with a planned release through the Anna developer platform.
Demo
The live demo is available while development continues.
Code
- Repository: github.com/udirobert/brume
- License: MIT
How I Built It
Brume is built with React, TypeScript, and Vite, with Progressive Web App support for the offline experience.
The main pieces are:
- MediaPipe for on-device camera and pose processing.
- Llama 4 Scout 17B via Cerebras for optional AI-powered personalised insights.
- FastAPI and Hetzner for the vision service.
- Supabase for optional persistence.
- Netlify for hosting.
- GitHub Actions and Dependabot for CI checks and dependency updates.
- GitHub Copilot in VS Code to help me iterate during development, including work on the breathing-pattern logic, pose pipeline, and vision service.
The architecture deliberately separates breathing from AI. The core experience does not depend on a model call, and the hosted AI companion requires a network connection. That means the app can still do its most important job even when the AI is unavailable.
Why Does Open Innovation Matter?
A calming tool should be easy to access, adaptable, and respectful of its users.
Open-source building blocks give me the freedom to inspect, modify, and replace parts of the stack as Brume evolves. MediaPipe enables on-device pose processing, while an open-weight model gives me more flexibility over how personalised coaching develops. The model is currently accessed through Cerebras, so AI coaching is not yet fully offline.
That distinction matters. I don't want to call a hosted AI experience fully local just because it uses an open-weight model.
The next step is to explore a smaller open-weight model for a genuinely offline Zen companion. I'd also like to experiment with locally generated ambient soundscapes, so Brume can become more personal without making a network connection essential to its core purpose.
My Agent Session
I’m using GitHub Copilot in VS Code as part of ongoing development, with inline suggestions, chat, and agent mode helping me iterate on Brume’s implementation.
The Lazy Payoff
The best outcome is a tool that asks very little of you.
When you have only a few minutes before something stressful, you shouldn't need to configure a complicated routine, create an account, or wait for an AI response just to take a breath. Open Brume, choose a pattern, and begin.
Sometimes the most useful technology is the technology that helps you pause — and then gets out of the way.
Prize Categories
Best Use of GitHub Copilot — Brume is being developed with GitHub Copilot in VS Code, including inline suggestions, chat, and agent mode. GitHub Actions also automates linting, typechecking, tests, and builds. Copilot-assisted development is the primary entry path; the CI workflow supports the project's development process.
Built for Hacktoberfest 2026 Week 3: Slow and Steady.
Top comments (0)