Skip to main content
The ThreadCat mark: a cat head holding an orange and teal swirl, two threads in one orbit.ThreadCatPinkPromise

Beta

ThinkPink (beta)

The governed assistant. Your machine, your record, your rules.

ThinkPink is an assistant that runs at home. It works with local models, the kind that live on your own computer, and it is governed by design: a built-in gate proposes, decides, checks, and keeps the receipts. Trust here is not a promise you have to take. It is a record you can read.

  • Local by default. The model runs on your machine. There is no account with us, because there is no us.
  • Governed by design. Its design law: a gate before every action, decisions filed with receipts, and one watched cloud lane at most, if you ever connect one at all.
  • Yours by design. Your context stays yours. The record does the remembering, and the record is a folder you own.
  • Free. Free for everyone. Nobody sells this packaging.
Read the source

Support runs through GitHub issues on the project repository.

New to local models? Start here.

First, the science, in one minute

An AI model is a big file full of numbers. When you "run" it, your computer reads that file and predicts, number by number, what text should come next. Nothing magical and nothing hidden: it is a file, and it runs on a processor.

Most AI you have used runs this file in a company's data center. You type, the words travel to them, their machine thinks, the answer travels back. A local model flips that: the file downloads to your computer, and the thinking happens where you sit. Your words never leave the building.

Why that matters

Your records stay yours. A local model has no server to leak, no account to breach, and no company to read along. What the assistant knows, it knows because a file on your disk says so.

It keeps working when the internet does not. No login, no subscription, no service outage between you and your own notes.

The footprint lands where you can see it. Training and hosting giant models in data centers takes a measurable toll: training GPT-3 directly evaporated an estimated 700,000 liters of clean freshwater, and AI's water withdrawal is projected at 4.2 to 6.6 billion cubic meters in 2027 (Li et al., Making AI Less "Thirsty", arXiv 2304.03271). Hosting general-purpose models costs far more energy per use than small focused ones; generality is paid for on every request (Luccioni, Jernite, Strubell, Power Hungry Processing, ACM FAccT 2024, arXiv 2311.16863). The energy cost of AI is old news, named in 2019 when the field was first made to count it (Strubell, Ganesh, McCallum, ACL 2019, arXiv 1906.02243), and researchers expect inference, not training, to dominate AI's electricity use as these tools spread (MIT News, January 2025). A small model on a machine you already own is the honest version of that math.

Setup, five steps, one cup of coffee

  1. Install Ollama. Go to ollama.com, download the installer for your Mac, run it. Ollama is the program that keeps model files and runs them for you.
  2. Get a small model. Open a terminal and type ollama pull gemma3:4b, then press enter. It downloads a few gigabytes once. This is the model the lab treats as its floor: small enough to run on an ordinary laptop.
  3. Get ThinkPink. Download it from the project's public repository (link in the flyer above) and follow the install steps in its README. Ad-hoc signed beta build: if macOS asks on first launch, right-click the app and choose Open.
  4. Open ThinkPink. It finds Ollama on its own. When you see the local-model indicator, you are connected to the model on your machine.
  5. Talk to it. Everything you type stays on your computer. The assistant's memory lives in files you can open and read.

What it cannot do (read this part too)

  • A 4-billion-parameter model is not a giant hosted one. It is smaller, slower to learn new things, and it will sometimes be wrong. Treat its output as a first draft from an eager intern, not a verdict.
  • ThinkPink is beta software. It was built to be tested in the open, and it says so on the tin.
  • The governance gate records decisions and keeps receipts. It is a discipline, not a conscience: a person still has to read them.

Sources

  • Li, Yang, Islam, Ren (2023, rev. 2025), Making AI Less "Thirsty", Communications of the ACM. arXiv 2304.03271
  • Luccioni, Jernite, Strubell (2024), Power Hungry Processing: Watts Driving the Cost of AI Deployment?, ACM FAccT 2024. arXiv 2311.16863
  • Strubell, Ganesh, McCallum (2019), Energy and Policy Considerations for Deep Learning in NLP, ACL 2019. arXiv 1906.02243
  • MIT News (2025), Explained: Generative AI's environmental impact.
  • Ollama, the installer used in this guide. ollama.com

Get ThinkPink

The download lives in the project's public repository, as a release built from that source.

Download from GitHub
  • Beta. A 127 MB zip for Apple Silicon Macs.
  • Ad-hoc signed, built by the lab from the repository's own source.
  • Requires Ollama plus a local model.
  • If macOS asks, right-click the app and choose Open.