About Low Key
Low Key is a calm way to buy a car. Don't get over-invested in the decision, and be confident you made a good one.
Pick a type of car and your ZIP, and set your own down payment, term and rate if you like. Low Key shows the new cars that hold their value best near you, what each really costs a month (the loan, fuel and sales tax), what it would cost to change your mind, and the lowest listed prices nearby, with where each price sits among the others listed near you and nationally. Bring a dealer's quote or a window sticker, and it checks the numbers and puts them next to its picks on the same terms. No account, no ads, and no money from dealers.
Why it works this way, and why there are no photos: why Low Key works the way it does.
The hackathon
An open-source entry to the Meta hackathon hosted by Cerebral Valley, Fall 2026. The code is public: github.com/ngoldbla/cv-meta-hackathon-2026.
How AI is used
Built with
Muse Code with Muse Wonder built the first phase of the app: every feature was written as a task file and built through the repository's task mailbox. Later tasks went through the same mailbox with a coding agent.
The Hairline skill drew the line-drawn cars; the impeccable distill method guided the interface clean-up.
At runtime
Muse Glimmer (30B, open weights) reads a photo of a dealer's sheet or a window sticker. It runs on Hugging Face Inference Providers (served by DeepInfra, which keeps no copy of the request), with our own serverless GPU as a backup. It copies the page and pairs each label with its value; plain code then finds the VIN, checks it, and picks out every number, and you confirm each one before anything is compared.
There is no chat. You choose a type of car and your terms with ordinary controls, and code does the rest: the cards, the charts and every figure are built by code, never by a language model. A number guard checks each figure on a card against the data that produced it. Cards and charts render with json-render from a fixed component catalog; the interface uses Iconiq UI.
In the weekly data pipeline
Timestral forecasts gas prices 52 weeks ahead on a serverless GPU that scales to zero: weekly, and again whenever EIA posts new prices.
Our belief
Car buying should be easy and suit your own needs. The open-source bundle makes it simple to roll your own with Muse and Vercel, so you can shop with your own preferences: luxury cars, motorcycles, other markets. See the roll-your-own guide.
Data sources
A licensed automotive listings and sold-vehicle data provider (listings, prices, and resale history).
U.S. Energy Information Administration (weekly gas prices).
Federal Reserve Bank of St. Louis, FRED (the 60-month new-car loan rate).
fueleconomy.gov (EPA combined MPG).
U.S. Census Bureau geocoder (your area from your location, when you tap “Use my location”).
State revenue and DMV pages and the Tax Foundation (sales tax).
iSeeCars (lifespan study).
Vercel (hosting, storage, and an approximate starting area). Hugging Face Inference Providers (photo reading). Modal (the reader's server and its backup GPU).
Open source used
The app's direct dependencies, generated from its package lockfile, plus the models and components above: Muse Glimmer (Apache-2.0), Timestral (code Apache-2.0, weights Modified MIT), Iconiq UI (MIT), vLLM (Apache-2.0).
@ai-sdk/openai-compatible 3.0.84 · Apache-2.0
@ai-sdk/react 4.0.166 · Apache-2.0
@base-ui/react 1.5.0 · MIT
@fontsource/ibm-plex-mono 5.3.0 · OFL-1.1
@fontsource/ibm-plex-sans 5.3.0 · OFL-1.1
@json-render/core 0.21.4 · Apache-2.0
@json-render/react 0.21.4 · Apache-2.0
@lucasmarkes/hairline 0.5.1 · MIT
@radix-ui/react-slot 1.3.3 · MIT
@vercel/analytics 2.0.1 · MIT
@vercel/blob 2.8.2 · Apache-2.0
ai 7.0.162 · Apache-2.0
class-variance-authority 0.7.1 · Apache-2.0
clsx 2.1.1 · MIT
lucide-react 0.564.0 · ISC
motion 12.42.0 · MIT
next 16.4.4 · MIT
react 19.3.1 · MIT
react-dom 19.3.1 · MIT
tailwind-merge 3.6.0 · MIT
tw-animate-css 1.4.0 · MIT
vaul 1.1.2 · MIT
zod 4.6.7 · MIT
Thanks
Fireworks, Hugging Face, Modal, Meta and Vercel, for their support in delivering this project.
Author
Dylan Goldblatt
ndgold@me.comComing next
Used cars one to five years old, single owner, no accidents.
True-shape drawings of each model, traced from photos with Meta's SAM 3.1.
Insurance and repairs in the monthly cost.