BOLT Outside 15: a tiny local Qwen planner that helps me close the laptop

Submission for the Hacktoberfest Open-Source AI Challenge — Week 1: Touch Grass.

What I Built

BOLT Outside 15 is a deliberately small, open-source CLI for moments when getting outside feels like another planning task. Pick observe, walk, or garden, plus a time budget from 10 to 60 minutes. A local open-weight AI model generates a short activity idea. The last line is not an "autonomous agent" action; it is an explicit reminder that the human decides whether the idea makes sense and conditions are safe.

I developed this standalone prototype in Catalonia, Spain on October 8, 2026, during this challenge's build window. It is not a rebranded release of my earlier BOLT software. I wanted to avoid turning a simple outdoor break into an account, mapping app, or more screen time.

The ideal interaction should last under a minute. Ask for one idea, leave the screen, and do the activity only if it is suitable.

Demo — real local inference

This command was executed on my MacBook with Ollama already running and Qwen2.5 1.5B installed:

node src.mjs --mode observe --minutes 15

An actual model response started:

Prepare: Choose a quiet spot and sit comfortably.

First Two Minutes: Breathe and observe colors, sounds and sensations.

Next Steps: Engage in a simple activity.

Stop or Alternative: Stop if uncomfortable and choose an indoor option.

This is a truthful excerpt, not a curated claim of perfect generation. The unfiltered live result also suggested talking to someone nearby despite my prompt discouraging approaching strangers. That exposed a limitation: a safety-aware prompt is not a guarantee. The CLI therefore warns that it is only a suggestion, not navigation, healthcare, accessibility or weather advice.

You can reproduce the demo with the commands in the README. It runs entirely on the local machine after downloading the model.

Code

Public MIT-licensed source, installation instructions and five automated tests:

ondmindmanagement-hub/bolt-outside-15

The repository was created for this challenge. It is a new and very small experiment, not a finished commercial product, and I do not claim adoption statistics or external beta users.

How I Built It

The pipeline intentionally keeps the authority boundary simple:

mode + minutes
      |
      v
deterministic option validation
      |
      v
bounded activity prompt
      |
      v
Ollama on localhost:11434
      |
      v
Qwen2.5 1.5B text generation
      |
      v
four-part suggestion + human approval reminder

I used Node.js, the built-in fetch API, Ollama, and the open-weight Qwen2.5:1.5b model. The model is central: it actually composes the suggestion. There is no prewritten activity database pretending to be an LLM. If inference fails, the program reports an error rather than inventing a successful AI reply.

The code deliberately never requests current GPS coordinates, a user account, API keys or health information. Input validation rejects unknown activities, fractional time budgets and durations above 60 minutes.

Verification on October 8:

• 5/5 automated tests passed using Node's built-in test runner.
• A separate real inference call completed against the locally installed Qwen model.
• The actual output's imperfection is documented in the README, rather than silently discarded.

AI assistance was used to draft and review this prototype and write-up; the code was also exercised with automated tests and real local inference. I reviewed the visible results and am not presenting the model as an independently safety-certified planner.

Why Open Innovation Matters

Local inference is meaningful here because a personal micro-break should not require telling a cloud service where I am, what I intend to do, or when I leave home. The actual model request stays on 127.0.0.1. The software and prompt are inspectable, and another builder can swap in a different Ollama-compatible model.

This is also a good example of why open does not mean perfect: users can run the tests, see the limitations, and improve the model-independent checks instead of trusting a black box.

What I Would Improve

Before I'd recommend this to strangers, I'd add a deterministic post-generation rule checker and an explicit consent screen; test the suggestions in multiple languages; use genuine user research to measure whether it helps anybody spend less time on screens; and build a smaller, accessible UI. Right now it's a working proof of concept, not an accessibility or safety product.

Prize Categories

I'm entering the overall Week 1 open-source AI challenge only. I have not integrated a partner product, so I am not claiming eligibility for sponsor-specific prizes.

Built by Omar Baró (Unfire), Spain. Source code: https://github.com/ondmindmanagement-hub/bolt-outside-15