One gentle action a week, planned for you by an AI that never leaves your laptop.
Photograph your vision board (or paste a few notes). A local Gemma model reads it and finds your goals. Choose one goal to focus on each week, then get one ten-minute action. Check in with Done, Too big, or Skip. Too big asks the agent to make the same action smaller.
Built for the DEV Hacktoberfest Open-Source AI Challenge, Week 3: Slow and Steady.
Teaser Video: SlothSteps Teaser
- Scan. Upload a board photo and/or paste notes. Gemma extracts the goals.
- Review. Edit or remove goals before saving. Saving a new scan replaces the old goals.
- Choose a weekly focus. Pick one saved goal; SlothSteps writes one concrete action of 1 to 10 minutes.
- Check in. Mark it done, skip it without pressure, or ask the agent to make it smaller.
- A gentle reminder. Choose a check-in day and download a recurring calendar event. Add it to your calendar to receive a reminder; it contains no goal or step details.
- Keep going. An unfinished focus carries over. After your check-in, the next weekly focus can be planned on the following week.
| Piece | What it does |
|---|---|
| Gemma (via Ollama) | Reads the board image and writes every step, fully local |
| Mastra agent | slothStepsAgent holds the rules for good steps and always returns typed objects (structured output) |
| Mastra tools | extract_goals, pick_next_step, shrink_step; each validates the model's answer and retries once with feedback |
| Mastra workflows | scanBoard, planWeek, shrinkStep decide the order of the steps; tools run directly as workflow steps |
| SQLite (better-sqlite3) | Goals, steps and run metadata, stored in ./data on your machine |
| Next.js + TypeScript | The two screens and the API routes |
The model is only asked for typed objects. The order of operations is decided by deterministic workflows rather than by the model choosing tool calls, which keeps a small local model reliable.
- Models run locally through Ollama. Your photo and goals are not sent to any server.
- The photo is held in memory for the request and is never written to disk.
- The calendar download contains a generic reminder and app URL only, never goal or step text. It is created only when you choose to download it.
- The database lives in
./dataand is gitignored, as issamples/. - The
runstable stores metadata only (duration, token counts, retries, error class), never goal text or images.
Prerequisites: Node.js 22.13 or newer, Ollama, Git.
git clone https://github.com/Namslay26/SlothSteps.git
cd SlothSteps
ollama pull gemma3:4b # any vision-capable Gemma works for photos
npm install
npm run dev # http://localhost:3000| Variable | Default | Purpose |
|---|---|---|
OLLAMA_MODEL |
gemma3:4b |
Which Ollama model to use |
OLLAMA_BASE_URL |
http://127.0.0.1:11434/api |
Point at Ollama on another host |
npm run extract -- samples/board.jpg # goals from a photo
npm run extract -- --text "run a 10k; learn guitar" # goals from text
npm run plan-next # plan this week's step if there isn't oneplan-next leaves an unfinished focus alone and plans at most one step per week, so it is safe to run every night. The calendar reminder is a separate, user-added event; the app does not send notifications itself.
Schedule npm run plan-next to run nightly.
# macOS / Linux: every night at 03:00
0 3 * * * cd /path/to/SlothSteps && npm run plan-next >> plan-next.log 2>&1On Windows, create a Task Scheduler task that runs npm run plan-next in the project folder.
src/app/ UI and API routes (goals, steps, plan)
src/mastra/agents/ the Mastra agent (Gemma via Ollama)
src/mastra/tools/ extract_goals, pick_next_step, shrink_step
src/mastra/workflows/ scanBoard, planWeek, shrinkStep
src/lib/ SQLite, validators, planner, metrics
scripts/ extract-goals.ts, plan-next.ts
MIT