100% OFFLINE LOCAL OLLAMA VISION + TEXT MODELS

Stop the rabbit hole. Keep the work.

The floating accountability companion that knows a coding tutorial is work, YouTube Shorts are not, and your screen never needs to become somebody else’s training data. Local Ollama models, one sticky note, zero cloud trackers.

Download open-source build
  • 0 bytes of screen data leave RAM
  • 30 sec to calibrate a workspace
  • $0 / mo API cost after install
focus67dumby.exe / 127.0.0.1 offline

TODAY'S FOCUS

Finish the release notes without opening a new rabbit hole.

67 min sprint 🟢 LOCKED IN • OLLAMA

09:41:08 VS Code + docs accepted

09:48:22 screenshot diff: 0.0%

09:52:10 no cloud call made

Live product preview / no account required

A sticky note that can tell “work” from “wander.”

Change the goal, model, and nudge state below. This is the same contract the desktop companion uses: a 512×288 in-memory frame, a foreground window title, and a local decision that never crosses 127.0.0.1:11434.

02 / observe

Floating sticky-note companion

gemma4:e2b VISION ON 127.0.0.1:11434
🔵 CALIBRATING WORKSPACE (30s)

Whitelisting VS Code, terminal, docs, and localhost for zero-friction switching.

Focus block progress 67%
67%
Preview state Manual override
Activity stream aria-live updates in the desktop app

30s calibration started — 4 workspace windows detected.

Pixel diff 0.0% — static screen costs zero GPU cycles.

Local model matched goal to React tutorial context.

This preview is a static contract for the desktop experience. A real session only evaluates the signals you enable.

Three ways to “focus”

Most tools force a trade-off. We refuse it.

You should not have to choose between blocking the tutorial you need and shipping your screen history to a remote tracker.

01BLOCK

Dumb website blockers

Cold Turkey / Freedom

Block all of YouTube indiscriminately. Great until the “distraction” is the coding tutorial, system design breakdown, or private stream you actually need.

Context-blind
02TRACK

Cloud AI trackers

Rize / Reclaim

Stream private window titles, client names, and screen activity to remote servers—then charge $15+/month forever for the privilege.

Data leaves home
03KEEP

Focus67Dumby

Local Ollama intelligence

Context-aware AI runs on your own machine at 127.0.0.1:11434. You get useful nudges, $0/month in API costs, and a screen that stays yours.

Context-aware + local

The engineering under the sticky note

Four small inventions that make local accountability feel human.

No black-box productivity score. The app composes tiny, inspectable signals into a nudge you can understand, tune, and turn off.

👁️ Hybrid vision + window title pipeline

gemma4:e2b or qwen3-vl:2b reads a 512×288 in-memory screenshot alongside the Win32 foreground window title. A 16:9 perceptual pixel-diff cache skips static screens below the <3% threshold, so a still editor costs zero extra GPU cycles.

512×288 frame + Win32 title → local verdict
  • Screenshot never written to disk
  • Static frame diff cache <3%
  • Deterministic fallback if Ollama is asleep

⏱️ 30-second workspace calibration

The first 30 seconds automatically whitelist every window you open—IDE, terminal, docs, localhost—for instant 0ms multi-window switching. Add popups with [+ Trust Window] and keep a rolling four-window breadcrumb trail for OAuth logins and handoffs.

VS Code›localhost›OAuth popup

🚨 Continuous escalating interference

Choose Gentle Toast, Toast + Sound, or Popup Takeover + Sound. On strikes 2, 4, 6… the sticky note forces itself frontmost, flashes red, and shakes—just enough to interrupt the spiral, never enough to hijack your machine.

1 / toast2 / shake4 / takeover

🔄 Zero-lock-in model switcher

Discover local models live through /api/tags. Swap from gemma3:270m at <400MB RAM to a multimodal vision model without a vendor account, backed by a deterministic offline heuristic fallback when Ollama is unavailable.

gemma3:270mgemma4:e2bgemma4:e4bqwen3-vl:2b
Your goal one outcome, in your words
Local context approved window and signals
Human-readable state explainable, reversible, yours

Why open & local innovation matters

The cloud is the wrong default for a screen-aware tool.

Open-weight models make a fairer accountability product possible: the sensitive loop stays on your machine, the economics stay honest, and the tool still works when the Wi-Fi disappears.

Browse the source on GitHub

Three reasons to keep inference close

  • Absolute data sovereigntySafe for engineers under NDAs: screenshots never leave RAM and window titles never leave localhost.
  • Unit economics that actually workPolling cloud vision every 10 seconds for 8 hours can cost $30+/month in tokens. Open weights make marginal inference $0.00.
  • True offline resilienceKeep a focus sprint on an airplane, a train, or unstable Wi-Fi; the local fallback still knows the boundary.
  • One-time purchase, not a meterPay for the polished desktop build once. There is no recurring API toll for having a private screen.

Pricing without the recurring guilt

Open source at $0. Polished at $19 once.

No subscription, no usage meter, no “your trial is over” nudge. Bring your own Ollama runtime or unlock the build that makes the first run painless.

Open-source / self-build

Bring your own Ollama

Clone, run, inspect, and customize the local companion yourself.

$0 forever

  • Clone from GitHub and bring your own models
  • python -m src.main or Docker
  • Deterministic offline heuristic fallback
  • Full source, no account, no API bill
Clone on GitHub
Most focused

Pro desktop edition

One click to locked in

The pre-compiled Windows desktop companion, ready for a real workday.

$19 lifetime / one-time

  • Focus67Dumby-Setup.exe / release bundle
  • 1-click Ollama setup helper
  • Custom alert packs + escalating nudges
  • Lifetime updates for one Windows device

Payment methods in the demo: UPI, Razorpay, cards, or Stripe. The checkout ships in simulation mode until you set a payment link or webhook in window.FOCUS67_CONFIG.

Setup / docs / bring your own runtime

Three commands. Then the note starts watching the boundary.

Ollama owns the model files. Focus67Dumby owns the tiny local loop around them. Use the standalone Windows build or run the open-source build in a Python or Docker environment.

WINDOWS 11 NATIVE GUI LINUX HEADLESS / DOCKER VERIFIED PYTHON 3.11+ OLLAMA LOCAL OPEN SOURCE
Terminal bash / local-only 
# pull the recommended local vision model
ollama pull gemma4:e2b

# start the open-source companion
python -m src.main

# or build the verified headless image
docker build -t focus67dumby .

✓ Ollama reachable at 127.0.0.1:11434
✓ screenshot cache armed at 512×288
→ focus state: calibrating workspace
One command to check. One goal to begin. See the contract