Focal

Pitch
Architecture
Agent Background
Agent: Focal Coordinator for Shopping
Simulation: Glasses Perspective
Simulation: Shopping for Airpods
User Interface on App

Inspiration

Quick: when did you last check your phone? Probably during a conversation, in a store, or halfway across a street. Every question ("is this a good price?", "what's the score?") pulls you into a screen and out of the moment.

Smart glasses could fix that, but only if they don't become another screen. So we built Focal around one idea: your agents work in the background, and your display stays clear until you want it. You ask once without looking down, a team of AI agents goes and does the work, and the answer appears in your line of sight only when it's ready. Then it's gone.

What it does

Focal is a team of AI agents that lives quietly behind your Even Realities G2 glasses.

Ask, then keep living. Say "Hey Focal" and ask anything. While agents search, compare and check, your lens stays out of the way. You get one short answer when it's ready.
Shop without the five tabs. Ask about a product and three agents work at once in the background: one finds trusted prices, one reads expert reviews, one finds a better alternative in your budget. A verifier checks every price. You see three picks, not fifty links.
Act only with your OK. One tap adds a pick to your Amazon cart, then Focal confirms it's there. It never checks out, and it never acts without your approval.
Any web page, made glanceable. Focal opens the best page for your question and rebuilds it into a few calm screens built for the lens. Suggestions for next steps wait until you've had time to read. Bookmark anything for later.
A display you control. "Hey Focal, hide the display" blanks the lens completely while Focal keeps listening, and anything new waits until you ask to see it. "Show the display" brings it back.
The rest of your day. Find a song or video, get a quick news brief, or draft a text that only sends once you approve it.

How we built it

Quiet on the surface (Even Hub SDK + Claude). The G2 shows 576×288 pixels in 16 shades of green, with no camera. Focal reads web pages, not your surroundings. A small server pulls the main content out of a page, and Claude turns it into a few screens with one idea each. Images become 4-bit greyscale so they read on a see-through lens. Voice runs on a wake phrase with StepFun speech-to-text, and you can navigate with the temple touchpad or the R1 ring.

Powerful underneath (Fetch.ai). 14 uAgents work in the background, each with one job: price, reviews, alternatives, a price verifier, memory, cart, search, YouTube, news, Spotify and a messenger. The FOCAL coordinator plans each request, sends specialists out in parallel, retries the ones that fail, has the verifier check prices, and combines the answers. Every agent speaks the Agent Chat Protocol and is registered on Agentverse, so ASI:One can find it. The full shopping flow also runs inside ASI:One with Interactive Cards, plus the Payment Protocol (Stripe test mode) for a paid price watch.

Calm control (FastAPI + Supabase + Paper). An orchestrator routes each request and streams real agent progress to the phone and dashboard, never to your lens. Supabase stores your sign-in, preferences (price, quality or balanced), history and every approval. We designed the landing page and dashboard in Paper first, with one green accent taken from the G2 display.

Real data, safe actions. Prices, reviews, news and videos come from public pages, with no paid data APIs. The cart runs in its own Chrome window and has no checkout code at all. We used Devin to write code too: it shipped fixes to our pitch page as a pull request we reviewed and merged.

Challenges we ran into

Saying less. A web page is built for a big colour screen. Deciding what to cut so each lens screen says one thing took many rounds.
Images on a see-through lens. Black is invisible on the G2, so normal photos glare or vanish. We convert every image so bright photos become line art and white backgrounds disappear.
Agents that don't embarrass you. Early on, Focal suggested a news headline and an AirPods case as "alternatives" to earbuds. We added rules to tell real products from headlines and accessories, and to match model numbers exactly.
Knowing when not to show something. Interrupting is easy. Holding popups while the display is hidden, and waiting before suggesting next steps, made Focal feel calm instead of busy.

Accomplishments that we're proud of

The work happens in the background, and the lens only lights up with the answer.
One spoken question goes all the way to a verified price and a confirmed cart.
Honest when unsure: if a price can't be confirmed, Focal says so, and nothing is sent or bought without your OK.
About 650 automated tests across the agents, backend and glasses app.

What we learned

On a lens, every word costs attention. The best screen is often no screen.
Small agents with one job beat one giant prompt. They're easier to test, retry and run in parallel.
Accuracy is the product. One wrong match costs more trust than a slow answer.
Agents can prepare things, but the final buy or send should always be a human tap.

What's next for Focal

Agents that learn your priorities and act on them automatically.
Quiet price-drop alerts that only appear when they matter.
Voice that works in loud stores and on the street.
Video summaries you can glance at instead of watch.

Tracks

Applied Intelligence: AI that saves you from overpaying and buying blind, without pulling you out of the moment.
Fetch.ai: a 14-agent team on Agentverse working in the background, fully usable in ASI:One with Interactive Cards and the Payment Protocol.
Even Realities: Quiet Tech by design. Agents work out of sight, the lens stays clear, and help appears only when you ask for it. Find it on Even Hub as Focal.
Paper: designed in Paper first, with one calm look shared from the dashboard to the lens.
Cognition: Devin wrote code for us as reviewed, merged pull requests.