reflect

A journal that thinks with you, and a finished example of what NobodyWho lets you build.

Building apps with local AI? NobodyWho is a young, genuinely promising engine for running language models on your users' own hardware: no API, no key, no server, no per-call cost. reflect is a complete app built entirely on it, running offline with the model loaded once and answering in under a second. If that is what you came for, go star NobodyWho ⭐.

One page a day. You write, and when you want, you click Ask me something and one open question appears in the margin, right where you stopped. Not "how did that make you feel," but something you actually skipped over.

It is deliberately not a chat. No transcript, no input box, no send. Your words stay where you typed them and you keep writing past the question. That one property is most of what separates a journal from a chatbot.

Demo

A 20-second walkthrough: one margin question a day, then your whole year read back. The preview is silent; watch with sound.

Powered by NobodyWho

Every question is generated locally by NobodyWho, an inference engine built on llama.cpp that runs the model on your own hardware (Metal on Apple Silicon, Vulkan on a GPU, CPU as a fallback).

It is the single choice that makes the rest possible: no server to trust means privacy is structural, not a policy. journal/engine.py is the only file that talks to it, so if you are building something with local AI, that one file is the whole integration. Star NobodyWho ⭐ and see how little it takes.

Privacy

The app runs entirely on your machine and has no way to send your writing anywhere. That is a property of how it is built, not a promise in a policy:

• No account, no cloud, no network calls, ever. Your words never leave the device.
• Your files stay yours, as plain Markdown in ~/Documents/journal/. Read, back up, grep, or delete them without this app.
• Proven, not claimed. test_no_network.py blocks every socket, then exercises storage, the model, and a headless window, so the no-upload promise fails loudly if any code ever reaches for the network.

Install

git clone https://github.com/nikolovv861/reflect && cd reflect
python -m venv .venv
source .venv/bin/activate        # Windows: .venv\Scripts\activate
pip install -e .
reflect

First launch downloads the model (~2.5GB, once). After that it opens offline in about two seconds. Run python download_model.py first to get the download and the one-time shader compile out of the way.

Requirements

• Windows, or macOS on Apple Silicon (Metal accelerated). Linux needs glibc 2.38+ (Ubuntu 24.04 or newer); the nobodywho 2.0.0 wheel fails to import on older versions.
• Python 3.10+, ~8GB RAM.
• A GPU is strongly recommended: sub-second questions on an RTX 2070, 15 to 30s on CPU alone.

What's inside

• Practices. Pick a frame (morning pages, what I'm avoiding, attention inventory, a decision you're sitting on) and the page opens with a real prompt. Each has an arc, not a streak. They are just Markdown files in journal/practices/, so adding your own takes no code.
• Review your year. Fourteen questions that shape a year, then Where you stand: a genuine reading drawn on the local model, not a mirror.
• Suggest a line (Ctrl+Space) offers one continuation in your own voice. Nothing is saved until you accept it.
• Your past, searchable. A sidebar of past days and standing notes; Ctrl+F searches everything.

Plain Markdown on disk, one file per day. Writing is never blocked by the AI: if the model is missing or busy, this is still a text editor that saves your work.

Tests

python -m pytest

140 tests, no model required: the engine runs against a stub, so the suite finishes in well under a second.