Week 1 of the Hacktoberfest Open-Source AI Challenge asked us to build something that gets people outside. Here's what I shipped: **GrassTrace* — a local LLM agent that reads your stress signal and generates a concrete 10-minute outdoor mission.*
If you spend your day in front of markets, dashboards, or a ticket queue, you know the feeling: eyes tired, mind buzzing, and "take a break" never converts into an actual action. Nudges don't work because they're generic. What actually works is a specific task: go to the street-corner park, find five different colors of flowers, listen for bird calls, feel the wind. Ten minutes. Done.
GrassTrace is that agent. It runs 100% locally — open-weight model, no API calls, your data never leaves your machine.
The problem: stress signals you can't act on
A stress signal like "I've been staring at charts for four hours" is real, but it doesn't tell you what to do. Generic advice ("take a walk!") is exactly what your brain ignores. The fix is to make the advice concrete, sensory, and short enough to not feel like a chore.
What GrassTrace does
You describe how you feel — stress type, intensity, how many minutes you have — and a local LLM generates a structured mission:
- diagnosis — one sentence on why you need to leave the screen
- mission — a walkable place, a specific action, and 3 observable details (colors, sounds, textures, scents)
- reflection — a guided question for when you come back (closes the loop, makes it stick)
Example output (Chinese, the model is bilingual):
{
"diagnosis": "眼睛疲劳可能是因为长时间盯着屏幕，建议休息一下并进行户外活动。",
"mission": {
"action": "在楼下人行道上走动，观察周围环境",
"place": "楼下人行道",
"observe": ["花的颜色", "树的形状", "路人的表情", "路旁的气味", "路面的温度"],
"duration_min": 10
},
"reflection": "10分钟的观察活动之后，你是否感觉到眼睛的酸胀有所缓解？"
}
Architecture: one inference pass, JSON in, JSON out
Three-step prompt pipeline in a single local inference pass:
- Diagnose — why leaving the screen matters right now
- Mission — action + walkable place + 3 specific observations + duration
- Reflection — a guided question for when you return
The model is constrained to output nothing but a JSON object; invalid outputs are retried with schema validation (max 3 attempts). A deterministic seed keeps output reproducible. The system prompt explicitly forbids market commentary or investment advice — this is not a financial tool, it tells you to go outside.
Stack:
| Piece | Choice |
|---|---|
| Inference | llama-cpp-python (OpenAI-compatible API, CPU) |
| Model | Qwen2.5-3B-Instruct GGUF Q4_K_M (open-weight, ~2 GB, bilingual) |
| UI | Streamlit with a "I'm back" check-in loop |
On my reference machine (i7-4790, 16 GB RAM, no GPU) a mission takes ~30 seconds. That's the honest cost of local CPU inference — and the privacy/ownership tradeoff is worth it.
Why local matters
- Privacy by default: stress is personal. No cloud, no logs, no prompts sent anywhere.
- Zero marginal cost: the model runs on hardware you already own; you can even disconnect the internet.
- The spirit of the challenge: open-weight models + local inference is exactly what this week's challenge asks for.
Try it
GitHub: https://github.com/fengyuGbt/grasstrace
The README has full setup: download the GGUF from ModelScope (or Hugging Face), pip install llama-cpp-python streamlit, one command to run, and you have a working agent.
Roadmap
- English mission cards (the model is bilingual; UI is currently Chinese-first)
- Auto-trigger on market volatility — the stress signal comes to you instead of you typing it
- Voice input and a PWA shell for phone use
AI disclosure: this post and the project were built with heavy AI assistance; the author reviewed, tested, and verified everything. The project itself is fully local — the AI that powers it runs on your own machine.
Top comments (0)