Stillbench

Live demo: stillbench.onrender.com · Integrations: /integrations · DEV submission (Week 1 — Touch Grass): draft on DEV · Repo: github.com/smriad/stillbench

Sit outside for twelve minutes. Stillbench picks a nearby bench, speaks a short Gemma brief once, then times your sit while the phone stays in your pocket. It is not a quest app, a streak game, or a garden planner—the screen is the short part; the sit is the product.

Optional This sit is for… names who the sit is for. TabPFN shows a plain why this bench line (light, last bail, leave probability). Take me outside is twelve minutes. The street is too much is six minutes at the closest bench (the roof counts). Sit with someone adds a line you can play for the person next to you. Check-in text shares a return time, not GPS. Afterward: one noticing prompt, optional debrief for the next visitor, and Copy sit story (markdown) on the done screen. Practice sit, twenty seconds is the same flow and is not training data.

Briefs can mention the air outside when STILLBENCH_HOME is set (Open-Meteo). The latest brief and voice file can stay on the phone for offline replay.

Demo (about sixty seconds)

On your phone: open the live demo → optional dedication → Take me outside → listen once → I'm leaving → I'm sitting → pocket the phone until the chime. Free tier may sleep; first load after idle can take ~1 minute. Check /integrations for partner status on production.

How it is built

Open-source AI is the spine:

1. Mastra workflow depart — read sky → TabPFN scores benches → pick bench → Gemma writes the brief → save walk → optional ElevenLabs voice.
2. Gemma (google/gemma-3-4b-it) via OpenRouter on Render, or Ollama locally (gemma3:4b). If inference fails, a built-in composer still gets you outside (src/lib/ollama-env.ts treats blank OLLAMA_API_KEY on remote hosts as unset).
3. TabPFN (Prior Labs) forecasts leave probability per bench from walk history and picks the farthest bench you are still likely to walk to.
4. Tinker (optional): half the briefs from a model tuned on sits that finished (npm run train:briefs).
5. SerpApi at build time adds real parks near STILLBENCH_HOME (coordinates are not sent per walk).
6. MongoDB Atlas on Render stores walks and integration health; Tiger Data logs sit events; Backboard may echo one debrief line into the next brief.
7. Sentry — @mastra/sentry with SENTRY_DSN; agent, TabPFN, and ElevenLabs spans (see docs/submission-screenshots/06-sentry-integrations.png).

flowchart LR
  tap[Take me outside] --> sky[read-sky]
  sky --> score[forecast-benches TabPFN]
  score --> pick[pick-bench]
  pick --> brief[write-brief Gemma]
  brief --> speak[ElevenLabs voice optional]
  speak --> timer[phone timer]
  timer --> note[one-line notice]
  note -.-> brief

Loading

Start in code: src/mastra/workflows/depart.ts, then src/lib/forecast.ts, src/lib/write-brief.ts, src/app/walk-app.tsx.

Why open innovation matters

Closed wellness apps could nag you to go outside. Stillbench needs open pieces to swap the brief model, train on your own bails, keep GPS off the server while personalizing distance, and run the departure agent on a harness you can read (Mastra).

Open piece	What it does
Gemma	Spoken departure brief on open weights
TabPFN	Bench choice from tabular walk history
Mastra	Five-step depart workflow
Tinker	Briefs tuned on sits that finished
MongoDB Atlas	Walks off ephemeral Render disk

ElevenLabs is optional closed voice on top; the open stack still works without it.

Real sit (Touch Grass)

Production walk: Banani Block C Park (SerpApi park in the pool), twelve minutes, dedication Hacktoberfest Week 1, debrief Generator hum from the club wall and someone washing a car two streets over. UI screenshots in docs/submission-screenshots/ (bench/timer from the same app the same evening; area photo CC BY 2.0 ASaber91).

Build log: DevRelay agent session #704 (curated; publish on DEV if you embed it in a post).

Run it locally

ollama pull gemma3:4b
npm install
cp .env.example .env.local
npm run dev

Open http://localhost:3000. Everything in .env.example beyond Ollama is optional. /integrations probes partners on load (on / set / broken / off). Integration health persists in MongoDB when MONGODB_URI is set.

Piece	Turn it on	What changes
Gemma	OLLAMA_BASE_URL, OLLAMA_MODEL, OLLAMA_API_KEY	Open-weight brief; composer fallback if inference fails
TabPFN	PRIORLABS_API_KEY	Bench choice from TabPFN instead of baseline
ElevenLabs	ELEVENLABS_API_KEY (text_to_speech, sound_generation, speech_to_text)	Brief voice, chime, spoken debrief
Mastra / Sentry	SENTRY_DSN	Workflow + agent traces
MongoDB Atlas	MONGODB_URI	Walks, attempts, integration health (required on Render)
SerpApi	SERPAPI_API_KEY, STILLBENCH_HOME, npm run benches:find	Real parks in the bench pool
Tinker	TINKER_API_KEY, TINKER_BRIEF_MODEL	~50% of briefs from tuned model
Backboard	BACKBOARD_API_KEY, BACKBOARD_ASSISTANT_ID	Debrief memory for next brief
Tiger Data	TIGER_DATABASE_URL	sit_events in Timescale
Temporal	TEMPORAL_ADDRESS, npm run temporal:dev, npm run worker	Durable server timer (optional)
Arduino	STILLBENCH_DEVICE_TOKEN	UNO R4 WiFi pebble in hardware/

What is saved

Each sit or bail appends to data/attempts.csv (or Atlas with MONGODB_URI). Rows: hour, distance, light, last bail, whether you sat. GPS is never read. Debrief audio is transcribed by ElevenLabs; only text is kept.

Deploy on Render

1. New → Blueprint → this repo (render.yaml, Node 22, /api/health).
2. Env vars: dashboard, or export RENDER_API_KEY=rnd_... then npx tsx scripts/sync-render-env.ts (copies from .env.local).
3. MONGODB_URI + IP allow list. SERPAPI_API_KEY + STILLBENCH_HOME for parks at build time.
4. OpenRouter: OLLAMA_BASE_URL=https://openrouter.ai/api/v1, OLLAMA_MODEL=google/gemma-3-4b-it, non-empty OLLAMA_API_KEY.
5. Temporal worker in render.yaml is commented out until Temporal Cloud + paid worker; phone keeps time by default.

Alternative Gemma host: deploy/digitalocean/cloud-init.yaml (Ollama behind bearer token on *.sslip.io).

Tinker training

npm run qualify:walks
npm run seed:walks      # optional, until 8 exportable sits
npm run train:briefs

Tests

npm test

Hacktoberfest 2026

Built for Week 1: Touch Grass (#hf26challenge, #devchallenge). Partner stack on the DEV write-up: Render, Gemma, TabPFN, Tinker, Mastra, MongoDB Atlas, SerpApi, ElevenLabs, Backboard, Tiger Data, Sentry Agent Tracing, Arduino (optional pebble).

Submission assets: docs/submission-screenshots/.