VR Apps

For this hackathon, I’m building PeekaPlate around a simple idea: ordering food from home can feel more personal, more visual, and easier to navigate. An agent to guide you. A meal you can explore before you choose. PeekaPlate. Your next meal, brought into view.

VR Apps

Shake it! Mushroom is a cozy WebXR mushroom-foraging RPG for Meta Quest. You shrink into a glowing acorn village and pick living mushrooms with your bare hands: grab the stem, wiggle it to loosen the roots, then pull. Haptics, soil crunches and a juicy pop sell the feel, and every species feels different (puffballs pop easily, fly agarics are stubborn). Your guide is Jev, a firefly companion powered by TypeSafe's Jev System One model through Vercel AI Gateway. About four times per second it reads the live game state (your hand, root looseness, quest, idle time) and decides in ~200-400 ms when to speak, which line fits, its mood, and which mushroom to fly to. Its decisions are shown in a small companion card (desktop) or on your left wrist (VR) without blocking play. Follow Jev's "PULL NOW!" call for a PERFECT pull and bonus coins. RPG loop: Grandma's quest, a collectible codex with rare variants, harvest grades, coins, and a Thorn Stick reward. Finishing the quest lifts the fog and opens the Grass Sea, a giant meadow with more mushrooms and a dandelion tower to explore. Built with Meta IWSDK + Three.js (WebXR, hand tracking with controller fallback, free locomotion with snap turn and teleport), TypeSafe Jev via Vercel AI Gateway in a Vercel serverless function, Tripo-generated 3D assets, WebAudio procedural SFX, deployed on Vercel.

VR Apps

VibeCheck is a mixed-reality dance and fitness coach built for Meta Quest 3. After the pandemic and the explosion of TikTok dances, a lot of people learn dance moves from online videos. A video shows you the move and a mirror shows you yourself, but neither tells you what you're doing wrong. With VibeCheck, you stand in front of a real mirror wearing the Quest and the headset's passthrough camera watches your reflection. VibeCheck estimates your posture on top of it, each limb is colored and labelled by how closely it matches the reference, with a ghost of the instructor's pose fitted to your body. The reference video plays beside the mirror and a voice coach gives you specific corrections. At the end of a routine, you get a report card showing your grade, with a per-limb breakdown and tips. Technologies: Meta Immersive Web SDK (WebXR immersive-ar with passthrough and hand tracking), the Quest passthrough camera through getUserMedia, MediaPipe Pose in the browser, Next.js on Vercel, Vercel Blob for videos and pose timelines, and the AI SDK through Vercel AI Gateway with fallback models: Llama 4 Scout writes the coach lines, Llama 4 Maverick writes the report card, OpenAI TTS speaks them. How it meets the VR Apps criteria: Mirror Coach is an immersive WebXR mixed-reality session that runs in Quest Browser from a live Vercel HTTPS URL, with nothing to install. We tested it end to end on a real Quest 3 in front of a mirror.

Agents & Factories

CASH is a shell that a team and its agents share. Sign in with a username, enter a room, open a task. On the left, the conversation: people and agents steer in the same stream, with commands like /agent, /approve, /pull and /status. On the right, the agent at work: a code file, a document, an image or a scheduled pipeline, with a viewer that adapts to the task type. Everyone sees the same state live from Neon Postgres. Each person gets a background assistant that reads what others did, explains it, and proposes the next action; reading needs no permission, acting does, so you approve with one click. Context is unified: /pull brings another task's artifact and discussion into the one you are in. Shell tasks are Unix-style agent pipelines (small commands joined by pipes) that keep running on a schedule after you walk away. This is how Xyne, Juspay's open-source org OS, thinks about agents: shared context, permission-aware, people and agents in one place.

VR Apps

Desktop Destroyer, but it's your real house. Quest passthrough + scene understanding turn your real walls, table, couch and windows into things you destroy with bare hands. Palm tool wheel (sledgehammer, chainsaw, flamethrower, machine gun, paint, flour bomb, fire ants, wasps, camel crickets), throw chairs through windows, raid the fridge, AI Gateway vision identifies unknown objects, DAMAGE $ ends in an insurance receipt. Built by the COGGS agent game harness in one 20-minute autonomous run.

VR Apps

Watching Formula 1 is hard. TV shows you whatever shot the director picks, and a grandstand seat gives you one car every 90 seconds. F1's latest fix is FPV drones that chase the cars at 300 km/h, but a drone is still someone else's camera. Trackside is a WebXR app for F1 fans that puts you inside the race. You can watch the real 2026 British Grand Prix from onboard any of the 22 cars, rebuilt from OpenF1 position, telemetry and team radio data. Or you can take the wheel and race the real field yourself. While you drive, an AI race engineer powered by Meta's Muse Spark talks to you over team radio, just like in a real F1 race. It calls out your braking and apexes against Antonelli's actual fastest lap, and it answers when you talk back.

Multi-modal Apps

30 Days is a purchase-permission agent: every buy asks first, and most buys wait a day. THE PROBLEM On a fixed budget, the expensive mistakes are made in the moment. I'm a CPA from Korea living in San Francisco until payday, and my job has always been to tell clients "not yet." 30 Days gives everyone that accountant, built into the Buy button. WHAT IT DOES Drop an ad (screenshot, link, or text) or just type what you want. The agent reads the ad's claims, finds the real product on Shopify through Shop MCP (photo and text search), pulls the real checkout total, compares it with similar listings, and decides by need: - Supplies (shampoo, detergent, a broken essential): price checked against the market median, approved now, single-use signed permit issued. - Gear (useful, not urgent): "Sleep on it." A durable Vercel Workflow sleeps 24 hours, rechecks the price and link, and your agent tells you when it's ready. Buy it tomorrow, if it's still true. - Treats (wants): never "buy now." 72-hour wait inside a small Treats envelope (10% of the budget), otherwise locked until payday. Everything stays inside the monthly budget. The checkout link (Shop continue_url) only appears after a permit is redeemed. THREE REAL CASES (try them in the app with "Try demo") 1. "My shampoo ran out" -> Pantene on Shopify, $10.00 vs $9.15 median -> approved, permit issued. 2. Allbirds ad screenshot -> Gemini reads the image, Shop MCP photo search matches the product -> Treats, locked, no buy button. 3. "Chelsea boots" -> Gear, not a hot deal -> waits a day, rechecked by a Workflow. MULTI-MODAL APPS TRACK CRITERIA - Two+ modalities: image (ad screenshots, Shop MCP photo search), links, and text. - Three models, three providers via Vercel AI Gateway: Gemini 2.5 Flash extracts, Claude Haiku 4.5 judges, GPT-4.1-mini audits; the stricter verdict wins. - Fallback: every role has an ordered chain. "Simulate model outage" fails the primary live and the routing panel shows the fallback model taking over. - Latency and cost: small fast models per role, per-call ms and cost shown in the UI; prices, medians, budgets and tiers are deterministic code, so models never do the math. AGENTIC An MCP server (/mcp) exposes judge_purchase, check_budget, request_permit, list_queue and recheck, with read vs spend scopes. Any agent can ask "can I buy this?" and get an evidence-backed answer; a read-only agent can never get a permit. Plug it into Meta's Muse: Muse Code takes it as a streamable_http MCP server in one settings entry ({"thirty-days": {"transport": "streamable_http", "url": "https://thirty-days.vercel.app/mcp"}}), so your Muse agent asks 30 Days before it buys. It works the same with any other MCP client. Next step: submitting 30 Days as a Muse connector so it shows up for every Muse user. BUILT WITH Next.js, AI SDK, Vercel AI Gateway, Vercel Workflows, Neon Postgres (budget, price history, permits, SF deal feed, SHA-256 hash-chained evidence ledger), Shopify Shop MCP. HONEST LIMITS Costs are estimates, thresholds are defaults, the ledger is tamper-evident (not a proof system), and the app can't block purchases made elsewhere. The demo's "Next day" button fast-forwards the Workflow wait.

VR Apps

Dreamspace is a multimodal world-building project: speak or type to Lumen, an in-world guide, and receive spoken replies plus changes to a shared 3D environment—summoned objects, portals, floating islands, and shifting atmospheres. It combines audio input, text interaction, audio output, and spatial/visual output, with VR as the interface. The local implementation uses Whisper speech-to-text, browser speech synthesis, Claude Code and local Ollama/Qwen model backends, with scripted fallback. Claude and Codex agents also cooperate in the world-building workflow: scene direction, visual scene generation, and integration into the spatial experience. The automated in-game handoff to Codex remains in progress. Team: yoshimis-sidegig. The live Vercel submission is the accessible spatial slice of that broader project: a Three.js/WebXR world with floating islands, portals, crystals, and five selectable atmospheres. It is based on the existing Dreamspace project. The local AI/voice backend and new IWSDK port are not connected to this hosted prototype, and AI Gateway is not yet integrated. The hosted scene and atmosphere controls have been verified on desktop; physical Quest testing is outstanding. No YouTube demo video is available.

VR Apps

Your real neighborhood as a live multiplayer world. People appear as characters at their real location on a map built from real streets, and local businesses advertise with tappable signs that show their products. On Meta Quest, walk the streets in VR or see people life-size in passthrough AR. Runs in the Quest browser from Vercel, no install. Built with Next.js, Neon + PostGIS, WebXR, and OpenStreetMap.

VR Apps

Venuelens is the augmented event space planner for Quest 3. Upload a venue blueprint, drag and drop your stage, tables, bar, and booths onto it, then step inside the plan in passthrough AR and walk the real room with your entire event at real scale before anything gets built. Plans travel as shareable links, so you go from laptop to headset in seconds. Built with Meta's Immersive Web SDK.

Agents & Factories

I built OpenCodex, a universal provider proxy for Codex and Claude Code (17.4k GitHub stars). Then I started running agents hard across many machines, and I became the bottleneck: dozens of Codex and Claude Code sessions, SSH into each Mac, and subscriptions split across Codex, Claude, Devin, Cursor and Vercel. GOAT Anywhere extends GOAT, our multi-agent desktop/server, to the web on Vercel. You type one instruction. The Main plans it with GPT-6 Astra on Vercel AI Gateway (with a fast fallback model) and routes up to five tasks in parallel to the right machine and subscription: your own Macs running GOAT (Codex, Claude Code) or private cloud machines on Vercel Sandbox, where OpenCodex runs API-key models through AI Gateway with no logins. Every task shows where it went and why, with live status, steer and stop, and a live fleet graph. Any machine's full GOAT opens with a one-time link. Factory jobs run on a schedule from Vercel Cron. Sign in with Vercel, and three private cloud machines start in parallel for you (about a minute; they stop on their own after 45 minutes). Subscription credentials stay on the machines; we never proxy them. Measured today (Oct 14, 2026): - One instruction → "5 tasks → 4 machines · planned by openai/gpt-6-astra · 9.2s" (Mac Studio, Travel MacBook, Home MacBook Pro, cloud machine); all five Done in about 37 s. - New Vercel account: 3 cloud machines created in the same second; "5 tasks → 3 machines · 9.0s", spread 2/2/1, all Done in 27 s. - Built in one day the same way: one Main GOAT thread dispatched 8 build threads to two home Macs; a browser agent set up Vercel; the Main merged and deployed.

Agents & Factories

How do you ai helps individuals improve how they work with AI and helps startups assess candidates’ AI proficiency through concrete evidence. We built a hosted dashboard with Codex and Claude Code plugins, automatic daily analysis, and interactive reports tracking interruptions, parallel sessions, longest uninterrupted tasks, and tokens per final output. Findings link back to session evidence, showing how someone frames problems, delegates, iterates, and verifies results across roles. Built with Next.js, Vercel, Eve, AI Gateway, Neon Postgres, and Clerk. Original transcripts stay local; bounded excerpts and metrics support the hosted analysis. I could not upload the video in social links but here it is: https://www.youtube.com/watch?v=U4geqXnAMPU Demo page with test data: https://hdyai-scorer-demo.vercel.app/reports/scorer-demo-2026-10-14

VR Apps

An immersive Texas Hold'em Poker game where you can play with 3 other AI players, You can go through all the rounds of a hand (pre-flop, flow, turn, river) and check/raise/call/fold at every turn as per the game rules.

VR Apps

Ready Player One Meta Quest 3

VR Apps

Spatial Merchant is a mixed-reality product showroom built with Next.js, React Three Fiber, Three.js, and WebXR. Alongside the Sora sculptural vase collection, it features protein and molecular design visualization, with potential applications in drug discovery research, education, and the development of models for 3D printing. A separate general showroom offers sample objects and a molecule-inspired collectible.

VR Apps

walkable city - stitched gaussian splat street segments

Agents & Factories

My CRM agent. I forward it a youtube sponsor lead inbound email, and it generates me a concept for the youtube video and drafts a reply. Here's my demo video: https://www.youtube.com/watch?v=5GRdmqQCF5M

VR Apps

Note to Self is a mixed reality app for Meta Quest: pinch the air, say what’s on your mind, and throw the note as a paper plane at a real surface. All in the palm of your hands, It’s for anyone who already leaves notes around the house and loses track of them.

VR Apps

VibeCheck is a mixed-reality dance and fitness coach built for Meta Quest 3. After the pandemic and the explosion of TikTok dances, a lot of people learn dance moves from online videos. A video shows you the move and a mirror shows you yourself, but neither tells you what you're doing wrong. With VibeCheck, you stand in front of a real mirror wearing the Quest and the headset's passthrough camera watches your reflection. VibeCheck estimates your posture on top of it, each limb is colored and labelled by how closely it matches the reference, with a ghost of the instructor's pose fitted to your body. The reference video plays beside the mirror and a voice coach gives you specific corrections. At the end of a routine, you get a report card showing your grade, with a per-limb breakdown and tips. Technologies: Meta Immersive Web SDK (WebXR immersive-ar with passthrough and hand tracking), the Quest passthrough camera through getUserMedia, MediaPipe Pose in the browser, Next.js on Vercel, Vercel Blob for videos and pose timelines, and the AI SDK through Vercel AI Gateway with fallback models: Llama 4 Scout writes the coach lines, Llama 4 Maverick writes the report card, OpenAI TTS speaks them.

Multi-modal Apps

When an expert leaves, their judgment leaves with them. Julia is a voice apprentice with a face. In Studio she watches an expert work in a live app, sees the screen, listens, and asks "why" at the moment a decision happens. She maps each decision and guardrail into a graph, then teaches the workflow to the next hire on a voice call, in their language. Multi-modal: voice in and out (ElevenLabs Agents, Scribe v2 Realtime), screen vision, and text. Claude Sonnet 5.5 runs through Vercel AI Gateway for vision, structured extraction and translation. A provider outage fails over from Anthropic to Vertex to Bedrock, then to Haiku 5.5. Voice falls back from WebRTC to WebSocket, and when the daily minute cap is hit, judges get the expert's recorded replay instead of a dead end. Every Claude call is priced on a live cost line. Built on Next.js, Neo4j and Vercel.

Multi-modal Apps

Drama Studio is a live dialogue workbench for writers and creators. Write or paste a multi-character scene, give each line a delivery cue, and hear it performed within seconds by Drama3, an open dialogue speech model running locally on a Mac with MLX. Audio plays while it generates, every version is kept for side-by-side comparison, and the final take downloads as a WAV. Optional: turn a picture or an idea into a script with a vision model via Vercel AI Gateway. The Drama3 model runs completely on one of our macbooks, on a forked version of SGLang-omni multi-modality inference engine with MLX Drama 3 support, where we implemented solely during the Hackathon. This will be a future Pull Requet of model support to the SGLang-omni project. (We are reviewers and active contributors of this community, and are responsible for this model support on Macbook anyways, so we did that during the hackathon as well) Still working MVP and model may go into problem of generating slient voices for certain length. Leave for future repair.

Agents & Factories

"Send a one-time code; an AI agent does the job on their Mac while you watch, then reports back." GigAgent lets you get work done on someone else’s Mac without a screen share. You write the job in a web dashboard and get a one-time code. The person at the computer pastes one command into Terminal (npx gigagent ABCD-EFGH), and an AI agent does the job on their Mac while you watch live. You can send it extra instructions or stop it at any time. When it finishes, you get a structured report on the dashboard and by Slack, webhook or email: what it did, the information you asked for, and key facts. Then the helper removes itself from their Mac. Typical uses are checking a new hire’s laptop is ready on day one, collecting system info for IT, diagnosing a slow Mac, and family tech support like tidying a Desktop or freeing disk space. A few design choices make it safe to hand to someone non-technical: The AI key never touches their machine; the server forwards every model call. Codes can be single-use and revoked instantly. The agent is told to work only inside one folder and to move files rather than delete them, logging every move so it can be undone. It runs on Vercel, with Neon Postgres for storage and models through Vercel’s AI Gateway (Claude Haiku 5.5 by default).

VR Apps

Dérive Home is a spatial memory map: a 3D scan of one apartment with time added. Switch between 1998, 2018 and 2026 and every object returns to where it stood; click one to see who owned it, its story, and the path it took through the rooms. Nine residents, including a couple who later divorced and a husband who died before the first era, are linked to the objects they left behind. The problem: every move-out wipes a home's memory. Property listings and 3D tours capture a space at one moment and can't hold what happened there. Dérive Home is for families keeping a home before it's sold, tenants and landlords handing one over, and historic houses that want visitors to add memories, not just read them. Five agents, one job each: Memory Capture ( turns a typed or spoken memory into structured data: time, object, place, people, event and emotion. It grounds the place to a real point in the scan and the year to the right era, then places the memory in the room as a new object. Voice Interview (talks with you about a memory and asks one question at a time to fill only the missing fields. If you don't know an answer, it records "unknown" rather than guessing. Historian ( links a memory to existing nodes across time (same object, moved, inherited, left behind, possible conflict) and refuses to merge when unsure. Literature Import turns a pasted story into scenes staged in the home. Temporal Compare describes what changed between two eras. A live activity panel shows which agent is working. The router calls one agent per task and never chains them for show, and each agent uses the model suited to its job.

Multi-modal Apps

Sellie: snap a photo, and an AI agent lists it everywhere it sells best, then takes it down everywhere else the moment it sells. demo link1: https://youtube.com/shorts/wWNxYnjX3Fc demo link2: https://www.loom.com/share/e17df555398c4667b319f074b198831e Sellie is for people clearing out a closet, not running a business. Reselling one jacket means hours of work: photos, research, pricing, writing a listing for each site, posting, and then remembering to delete the other listings when it sells. Most people give up and donate it. With Sellie you take a few phone photos and type the least you’d want to keep. Sellie handles the rest: Identifies the item from the photos: brand, model, size, condition and value. It reads tags and points out flaws. Makes listing photos. Your real photos are corrected and cut out onto a studio background, and labeled AI product and model shots are added. A second model checks every edited image against your originals for color, details and wear before any buyer sees it. Chooses where to sell. Clothes go to Poshmark and Vinted. Bulky or electronic items go to Facebook Marketplace for local pickup. Prices each marketplace backward from your floor, so after fees, shipping and haggling you still keep what you asked for. Writes one listing per platform in that platform’s style and within its limits. Posts for you. Claude reads Sellie’s posting kit over MCP and fills in each marketplace’s form in your own signed-in browser with Claude in Chrome. The item also goes live on your Shopify store. Stops double sales. Shopify is the inventory record. When the item sells anywhere, its quantity drops to 0 and every other listing is flagged for takedown. Shopify checkouts arrive instantly by webhook. Marketplace sales are found by a check that runs every 10 minutes and reads each listing’s public “sold out” status. Built with: Shopify (products, media, metafields, and the inventory lock), Vercel (hosting, Cron, and AI Gateway with automatic model and provider fallback), Claude Opus, Sonnet and Haiku 5.5 for vision and text, Gemini / Nano Banana for images, Neon Postgres, and an MCP server so any AI assistant can drive it.