For years, AI agents were stuck in chat windows. Then we opened the phone book.
Hosted by AgentPhone. 91 teams shipped agents that pick up the phone, work the inbox, and act in the real world.
Gavel sells your stuff for you. Text Gavel what you want to sell. It posts the listing, talks to interested buyers, negotiates the price, and schedules pickup. You hear from it once — when someone's on the way to grab the item.
Adam Zinebi, Drishti Jain, Thomas Kiefer
Built on Google DeepMind / Browser Use / AgentMail / AgentPhone
A real phone number you can call from a police station. An AI voice agent picks up, takes a 2-field intake (name + criminal charge), semantically routes the case to local criminal defense firms, sends real outbound intake emails on the caller's behalf, and remembers the case across calls. So the next time you dial in, it greets you by name and surfaces any attorney replies that came back. Every year, roughly ten million people are booked into U.S. jails. At any given time, nearly 70% of the jail population hasn't been convicted of anything - hundreds of thousands of people sitting in custody, waiting. The law guarantees them a phone call. Most of them don't know a criminal defense lawyer's number. The fallback is the public defender's office - overworked, underpaid, and not always available. When someone is booked at 2 AM, the PD's number rings to voicemail. A pilot in Santa Clara County found that people who actually spoke to a lawyer right after arrest were 28 percentage points more likely to be released pretrial, and spent 6 days in jail on average instead of 29. The difference between those two outcomes is a single phone call that gets answered. JailCall is the number that always answers.
Souvik Banerjee
Built on Google DeepMind / Browser Use / Moss / AgentMail / Supermemory / AgentPhone
Chief is an overhearing agent, an AI that listens in on a live conversation it isn't part of, holds context from every past conversation, and knows when to stay quiet and when one piece of information causes it to chime in without you ever breaking the conversation to look something up. This builds directly on "overhearing agents," a paradigm formally defined in recent research by Dr. Chris Callison-Burch's lab at UPenn. Dr. Callison-Burch is our research mentor, and Chief is the first working implementation of the agent class his paper describes. We built it as a customizable framework, not a single app: the same engine can be pointed at an investor meeting, a doctor's consult, or a sales call by changing what it watches for and which tools it has access to. Chief runs locally as a desktop app and never joins the call as a participant, so the people in the room don't see a bot. It captures audio on-device, transcribes with ElevenLabs Scribe, and resolves who is speaking before any model sees an utterance. At conversation start it resolves the counterparty and creates a frozen briefing of everything it remembers about them, with durable memory stored in Supermemory and low-latency semantic retrieval handled by Moss. When something worth saying comes up, a compact orb surfaces one sentence and then collapses back into a listening state. In the background it can also prepare real actions, drafting a follow-up email through AgentMail, running deeper web research through Browser Use, pulling a document, checking a calendar, or placing an outbound side call through AgentPhone, with anything irreversible held back for explicit approval. The core engineering challenge is latency on the judgment layer between a live transcript and the user's attention. Everything expensive, belief updates, memory retrieval, claim extraction, action planning, runs continuously in background threads, so the only thing on the hot path is a single cheap end-of-utterance decision that reads warm state rather than building it. The decision call has a hard timeout and fails to silence, because a missed whisper is recoverable and a wrongly-timed one is not. To keep false positives near zero, the decision agent is constrained to return either silence or one sentence, with speaker identity resolved locally before any model sees an utterance, and is tuned against a fixed suite of decoy utterances treated as a regression gate on every change.
Benjamin Liu, Jishnuu Senthil Kumar
Built on Browser Use / Moss / AgentMail / Supermemory / AgentPhone
Picks from each sponsor track.
We have no audio in my video demo. We just realized this. But I have the app deployed here: https://domainsdr.vercel.app/ DomainSDR is an autonomous domain broker agent. A domain owner enters a domain, price, floor price, and rules. The agent then works on its own: it researches likely buyers, finds contacts, sends low-volume personalized outreach using AgentMail, handles replies, negotiates within the seller’s limits, collects a Stripe deposit when a buyer is ready, and uses AgentPhone to schedule a post-deposit handoff call. The goal is to make domain sales less illiquid by turning each owned domain into an active outbound sales campaign, without requiring the owner to manually research buyers, write emails, follow up, or negotiate. It uses: - **AgentMail** for outbound and reply handling - **AgentPhone** for post-deposit calls and scheduling - **Stripe** for deposit collection - **Gemini** for buyer research, scoring, and reply understanding - **Apify / Browser Use** for buyer and contact discovery - **Supermemory** for long-term campaign memory and learned buyer/contact patterns - **Neon Postgres** for campaign state and transaction records In one sentence: **DomainSDR gives every domain owner an AI broker that keeps working until the domain gets a serious buyer, deposit, or handoff call.**
Carl Okpala, Jiyun Kim
Built on Google DeepMind / Stripe / Browser Use / AgentMail / AgentPhone
Riley is a single phone number that both parents and kids text. Parents check grades, monitor browsing, manage calendars, and approve spending. Kids request money and relay messages. Everything happens over iMessage ~ no app, no login, no screen. Just text.
Jacob Fu, Sebastian Jurado, Gabriel Dennis Shores
Built on Browser Use / Supermemory / Sponge / AgentPhone
AI Voice Agent that sits on your website 24/7 to call and bring back all the abandon checkout. It gets implemented with one line of script tag.
Sohazur Islam, Shanzila Ahmed, Maryam Alabbar
Built on Moss / Supermemory
GOFER is an autonomous errand operating system. Tell it what you need done, book a dentist, dispute a charge, order food and it dispatches a team of specialist agents that make real phone calls, browse authenticated sites with your profile, send emails, and prepare payments, all running in parallel. It stops before any irreversible action and asks for your approval. Every completed errand trains its memory so the next one is faster.
Carl Okpala, Jiyun Kim, Godson Ajodo, Elijah
Built on Google DeepMind / Browser Use / Moss / AgentMail / Supermemory / AgentPhone
Virtual host for restaurants that answers the phone and takes reservations
Stepan Arsentjev
Built on Stripe / Moss
David is your personal accountability coach, calls & checks in with you every day, manages your tasks on Notion, send money to organization you hate if you follow hit goals, books things for you that you've been putting off using the budget you give him
Ahmet Ilten
Built on Google DeepMind / Browser Use / Sponge / AgentPhone
Haggle is an AI concierge that negotiates local services for you over the phone. Text Haggle what you need ("car detail in SF under $100 this weekend") and it handles everything: - Finds providers (Browser Use) - Calls them live (AgentPhone) - Negotiates pricing within your budget - Books, pays, and sends a follow-up once you confirm (AgentMail, Sponge) All via text/ a custom UI. How it works: 1. You text Haggle Send a request with your service, budget, timing, and preferences. 2. Haggle researches providers Using BrowserUse, Haggle scrapes and enriches local provider data, then ranks providers using historical interaction data from Supermemory. 3. Haggle calls providers Haggle places parallel outbound calls via AgentPhone. During the call, a voice agent runs a real-time negotiation loop: - Gemini generates one turn at a time - Waits for the provider’s response - Counters if pricing is over budget - Confirms price, timing, and payment method If nobody answers, Haggle falls back to cold email via AgentMail. 4. You review options Haggle presents providers ranked by: - Price - Availability - Reliability - Historical interactions 5. Booking + escrow You choose a provider and fund escrow using: - Card - USDC Powered by Sponge. The provider receives a payout link, and funds are released once you confirm the job is complete. The sticky loop: With Supermemory, Haggle remembers: - Preferred providers - Scheduling habits - Budget preferences - Communication style - Historical pricing across bookings It proactively suggests re-bookings when recurring services are overdue. The more you use Haggle, the less information you need to provide. Stack: - Next.js 15 on Vercel (serverless, ~120s max duration for long workflows) - Convex for real-time state (jobs, leads, calls, escrow, chat) - Gemini Flash for triage, intent parsing, negotiation summarization - AgentPhone for outbound voice calls + iMessage/SMS - AgentMail for email fallback + follow-ups - BrowserUse for provider scraping + enrichment - Supermemory for persistent user context (preferred providers, scheduling habits, price history, communication style) - Sponge for USDC escrow, card onramp, virtual card payouts, ACH via Bridge
Rishi Thakkar, Muzart Tuman, Evan Taylor
Built on Google DeepMind / Browser Use / AgentMail / Supermemory / Sponge / AgentPhone
Antar is the agent a field technician calls from a basement. (formal tagline is AI for the trades) Skilled-trade technicians (HVAC, electrical, plumbing) do their hardest work in places phones don't get signal: mechanical rooms, basements, attics, crawl spaces. They're standing in front of a furnace at 9pm with a fault code they don't recognize, and the answer is buried in a 400-page manual on a laptop in the truck. Antar is the senior tech they can phone instead. How it works in the real world: Multimodal agent powered by Gemini. Point the phone camera at the unit, Gemini identifies the equipment, surfaces the fault code, pulls the wiring schematic & handles real-time voice and reasoning. AgentPhone for the field. Custom webhook routes a phone call to the same Gemini Live agent, grounded in OEM service manual context. The tech describes what they're seeing; Antar walks them through the diagnostic. Supermemory keeps the agent sharp. Persistent cross-session memory of the tech: their truck stock, the brands they service, the units they've worked on. Every call gets smarter. AWS handles the corpus. ~2TB of OEM service manuals on S3, retrieval pipeline running on EC2, feeding the agent during live calls. Transcript waits in the app. The moment the tech is back on wifi, the full call transcript, diagnostic steps, and referenced manual pages are sitting in their job log. Why this matters? A real technician, on a real service call, with no laptop and no internet, gets real work done by calling an agent. The agent acts in the physical world by directing the human standing in front of the equipment.
Harsh Mehta
Built on Google DeepMind / AWS / Supermemory / AgentPhone
Leakly is an autonomous property maintenance dispatcher that handles the entire repair lifecycle without human intervention. When a tenant reports an issue via chat or email, AI classifies the problem, selects and contacts the best vendor, schedules the appointment based on the tenant's availability, and pays the vendor in USDC on-chain via Sponge. Every step — from intake to payment receipt — is automated, with the tenant and property manager able to track progress in real-time.
Arjun Chidambaram, Dante Kakhadze
Built on Google DeepMind / Browser Use / AgentMail / Supermemory / Sponge
88 submissions, in alphabetical order.
Whether on the go or at your computer, talk to your personal AI that acts for you fixing prod or getting a pitch deck ready by giving you generated UI output + voice instead of text.
Alex Aridgides
Built on Google DeepMind / Browser Use / AgentMail / Supermemory / Sponge / AgentPhone
AgentOG is a human approval and execution-control layer for high-impact AI agent actions. It sits between AI agents and real-world systems like payments, bookings, forms, APIs, and external workflows. Before an agent executes an action, AgentOG creates an action fingerprint, gets human approval, issues a short-lived action-bound token, and blocks execution if the final payload changes.
Rahul Marri
Built on Google DeepMind / Stripe / Browser Use / Moss / AgentMail / Supermemory / AgentPhone
WordPress still powers 43% of the internet. Millions of those users are small businesses owners who don't want to change what's been working for 20 years and aren't tech savvy enough to use Claude Code. agentpress lets them make changes by calling/texting an agent that knows everything about their site and can verify success using a real browser, offering a preview URL before deploying to prod.
Troy Gunawardene
Built on Google DeepMind / Browser Use / AWS / Moss / Supermemory / AgentPhone
Calling your representative in the United States is the best way of getting your voice heard to address a problem. However, calls rarely ever happen because people don't have time or they don't want to get on a 1 on 1 call with a staff member. In this hackathon, we used AgentPhone to enable any user to use agents call on their behalf on issues they are passionate about. I used this system to tackle a problem personal to me: environmental destruction as a result of sewage in my hometown.
Joshua Chung, Vivek Shukla
Built on Browser Use / Moss / AgentMail / AgentPhone
AI-Native Go Fund Me. Write text about cause. AI generated fundraise page. Upload contacts. Browser use figures out info about contacts [and their connections, and their connections (2nd and 3rd degree)], their lives, what they've faced and if they're likely to donate and how much if so -> pushes it to a knowledge graph so in the future it can just use this as a data source -> AgentPhone/AgentMail to contact the donor -> stripe for accepting payments -> supermemory for keeping conversations with donors and future donor outreach
Aparajith Raghuvir, Vivek Chandrasekar
Built on Google DeepMind / Stripe / Browser Use / AgentMail / Supermemory / AgentPhone
Relocate — one phone number, a swarm of 17 AI agents that handles every task a US move requires, producing a verifiable real-world artifact (confirmation number, certified-mail tracking, AgentMail message ID) within 90 seconds. Built on PAVO, our peer-reviewed routing layer (TMLR 2026, 50K-turn benchmark) that runs cheap turns locally on Apple Silicon and escalates to cloud models only when needed. It is 25% cheaper, 34% faster, 71% less energy, 7.9x fewer failures than fixed-cloud. Just call +1 (618) 414-9537.
NarasingaMoorthy VeiluKanthaPerumal, Pranav Guhathakurta
Built on Google DeepMind / Stripe / Browser Use / AWS / Moss / AgentMail / Supermemory / Sponge / AgentPhone
Antar is the agent a field technician calls from a basement. (formal tagline is AI for the trades) Skilled-trade technicians (HVAC, electrical, plumbing) do their hardest work in places phones don't get signal: mechanical rooms, basements, attics, crawl spaces. They're standing in front of a furnace at 9pm with a fault code they don't recognize, and the answer is buried in a 400-page manual on a laptop in the truck. Antar is the senior tech they can phone instead. How it works in the real world: Multimodal agent powered by Gemini. Point the phone camera at the unit, Gemini identifies the equipment, surfaces the fault code, pulls the wiring schematic & handles real-time voice and reasoning. AgentPhone for the field. Custom webhook routes a phone call to the same Gemini Live agent, grounded in OEM service manual context. The tech describes what they're seeing; Antar walks them through the diagnostic. Supermemory keeps the agent sharp. Persistent cross-session memory of the tech: their truck stock, the brands they service, the units they've worked on. Every call gets smarter. AWS handles the corpus. ~2TB of OEM service manuals on S3, retrieval pipeline running on EC2, feeding the agent during live calls. Transcript waits in the app. The moment the tech is back on wifi, the full call transcript, diagnostic steps, and referenced manual pages are sitting in their job log. Why this matters? A real technician, on a real service call, with no laptop and no internet, gets real work done by calling an agent. The agent acts in the physical world by directing the human standing in front of the equipment.
Harsh Mehta
Built on Google DeepMind / AWS / Supermemory / AgentPhone
Apex is an AI receptionist for local businesses that can receive inbound calls, cold-call leads, qualify them live, handle objections with real-time product context, and close the loop with follow-up email, booking, and payment. It calls and response using AgentPhone and Gemini LIVE API, uses Moss to answer product questions mid-call, sends personalized follow-ups through AgentMail, collects payment through Stripe when the lead is ready, and uses super memory so every future call or email remembers the full customer history. It’s simple for any local business to use: upload your leads or customer list, connect your business details, and let Apex handle the calls, follow-ups, bookings, and payments automatically.
Thien Phu Quach
Built on Google DeepMind / Stripe / Moss / AgentMail / Supermemory / AgentPhone
Artificial People that each have an email address, phone number, and bank account. and a memory. They are able to take human actions on the internet, such as order doordash using their email, phone number and bank account.
Zuri Obozuwa
Built on Google DeepMind / Browser Use / AgentMail / Supermemory / Sponge / AgentPhone
An AI agent that brokers a used-car deal end to end for a Spanish-speaking buyer: it takes the inbound Spanish call, texts the human broker for one-tap approval, calls the dealer to price the car, places a deposit to hold it, calls the bank for financing, and calls the customer back in Spanish to confirm.
Puneet Pandian
Built on Sponge / AgentPhone
Ava is an end-to-end AI recruiting agent that founders can call. Ava understands the hiring requirements over a phone call, sources and ranks candidates, automatically calls and screens them, evaluates fit, and sends follow-ups with booking links to qualified matches, turning a founder conversation into an interview pipeline autonomously.
Anish Kallepalli
Built on Browser Use / AgentMail / AgentPhone / Google DeepMind
AI inquiry copilot for Chinese B2B exporters — LLM + telephony + CRM, with REST API & MCP server
Jinwei Han
Built on Google DeepMind / Browser Use / Supermemory / AgentPhone
AI video editing agent: Agent is triggered via text using agent phone. Analyzes videos at a singular frame level (leveraging Google Gemini multimodal models) to create sharper video edits. Provides progress updates using text and emails, using agent phone and agent mail.
Udayan Maurya
Built on Google DeepMind / AgentMail / AgentPhone
Banana Taxi is an AI agent layer for robotaxis that uses live vision, routing state, and voice calls to proactively coordinate with passengers. It lets a robotaxi explain where it is, what it sees, and what it is doing next, from landmark-aware pickups to reroutes down the road.
Vikram Kommera, Tarun Rajesh
Built on Google DeepMind / AgentPhone
Better Call Ron is your AI legal fixer: One phone number, zero hold music. Whether you're cuffed in a holding cell or shopping for a divorce attorney, Ron picks up, reads the room in under 30 seconds, and either cold-transfers you to a vetted lawyer in 90 seconds flat or walks you through the legal landscape with ease and helps you find legal resources. Check it out: https://better-call-ron-hackathon.vercel.app Core Features: - Lawyer Matching based on urgency, location and your budget - Memory context identifies the user from their phone numbers and retains their past info - Only escalates to a real lawyer in case of an emergency or if user explicitly asks (Ron tries to handle mostly everything on its own) - In urgent situations, Ron notifies your emergency contacts
Umair Usmani, Haris Jalal, Ibrahim Kamal
Built on Google DeepMind / Browser Use / AWS / AgentMail / AgentPhone
Callan is an agentic cold-calling web agency. It finds weak-online-presence local businesses, researches them, places sales calls, invoices interested customers, tracks payment, and kicks off website fulfillment through Lovable or v0. The app is a React/Vite operator console backed by an Express/SQLite server. It visualizes the whole agent workflow: lead discovery, Browser Use research sessions, Gemini structured reasoning, Supermemory customer memory, Moss call-time retrieval, AgentPhone calls, AgentMail follow-ups, Stripe invoices, website builds, and post-delivery growth recommendations. Mock-mode scripts let the full lifecycle run safely without live provider side effects.
Aditya Das, Rishabh Bansal
Built on Google DeepMind / Stripe / Browser Use / AgentMail / Supermemory / AgentPhone
It uses an agent to analyze your inbox and call you if there is an emergency. In the phone call you can also draft and send the email.
Abhiram Segu
Built on Browser Use / AgentMail / AgentPhone
voice agent help companies and organizations to book appointments and send email confirmation
Chao Xu
Built on Google DeepMind / Moss / AgentMail / Supermemory / AgentPhone
Chasm builds websites with stripe checkout in seconds over phone and text
Chaira, Saran, Michael, Aditya
Built on Google DeepMind / Stripe / AWS / Moss / Supermemory / Sponge / AgentPhone
Today's customer service AI is stateless and powerless — it doesn't know who you are, what you ordered, or where your driver is, and even if it did, it can't actually do anything about it. We're building a voice and SMS agent that picks up when humans can't (the driver mid-traffic, the support line on a 45-minute hold) and has full live context on the customer, the order, and the company's systems. More importantly, it can take real action mid-call — rerouting a driver's GPS, swapping an item in a grocery order, upgrading a flight — not just route you to a human queue. Call our number when you're at the wrong SFO terminal, and the agent answers for the driver, updates the destination on the driver's map in real time, and confirms the change with you in under a minute. It's the difference between AI that talks to you and AI that handles it.
Iminabo Roberts, Holy Agyei
Built on Google DeepMind / Browser Use / AWS / Moss / Supermemory / AgentPhone
“Indexing the Web for Agents” COSMIC Browser Memory is a web traversal memory layer for browser agents. Search engines indexed web pages for humans. COSMIC indexes how agents move through websites: page states, screenshots, visual anchors, actions, coordinates, failures, fixes, and successful workflows. It runs on top of cosmic-browser-use, our vision-based browser automation agent that can see and operate real websites. On the first run, the agent completes a task normally while COSMIC records the traversal path. After the task succeeds, COSMIC compresses that run into a reusable workflow, stores the executable route locally, and writes semantic workflow memory to Supermemory. On future runs, the agent retrieves relevant prior workflows, loads the executable route, adapts task-specific values, replays safe known actions, and hands control back to live vision whenever the website has changed or confidence drops. The result is a browser agent that does not rediscover the same websites from scratch every time. Every successful run becomes reusable intelligence for future tasks. COSMIC also integrates with AgentPhone and AgentMail. Users can call the agent live and ask it to browse, or email it a web task like: “Get me the latest 1040NR tax form.” The agent browses, completes the task, and replies with the result or attachment. Benchmarks: Task: ""Get the YouTube video description for the official Claude Code Launch Video"" Browser-use with Browser-use 1.0: "38.59s elapsed (done)"; Cosmic-browser-use: "34.80s"
Praveen Raj Uma Maheswari Shyam Sundar
Built on Google DeepMind / Browser Use / AWS / AgentMail / Supermemory / AgentPhone
Small event businesses often win jobs through messy texts, emails, and web requests, but fulfilling those jobs still depends on the owner manually coordinating people, supplies, invoices, and pay. The painful part is that every urgent request forces the owner to remember who is reliable, who is nearby, who can do each role, who has worked similar events, and who responds fast enough. At the same time, they still need to buy inventory, invoice the client, confirm the crew, and avoid releasing worker pay before the job is verified. When any step slips, the business risks unfilled shifts, duplicate outreach, late supplies, delayed billing, or unsafe contractor payouts. CrewLoop solves this by turning one event request into a closed-loop AI dispatch workflow that plans the job, shortlists contractors, gets owner approval, contacts workers, buys supplies, invoices the client, and sets conditional worker pay. Moss powers contractor memory such as skills, rates, reliability, and past-job fit, while Supermemory gives CrewLoop long-term operational memory across chats, events, vendors, preferences, and decisions. AgentPhone is required for live SMS and urgent calls, AgentMail is required for owner summaries, client invoices, and contractor confirmations, and Browser Use is required to source or purchase inventory through real web workflows. Stripe/MPP powers the invoice and payment-status flow, Sponge powers programmable payment holds and release rules, and together the sponsors make CrewLoop a real AI operations dispatcher rather than a static planning dashboard.
Ayush Ojha, Kush Ise, Dhrumil Shah
Built on Google DeepMind / Stripe / Browser Use / Moss / AgentMail / Supermemory / Sponge / AgentPhone
AI native biospecimen procurement platform for biotechCrovi uses agents to do the work of sample sourcing, audit. Teams go from prompt to a zip file in weeks instead of months.
kamil seghrouchni
Built on Browser Use / AgentMail / Supermemory / Sponge / AgentPhone
Deardiary — pick up the phone (it calls you), journal out loud, and the agent remembers. Over time it acts like a friend who listens and pays attention, surfacing patterns from past calls and quietly assembling a "Life OS" dashboard of mini-apps from what comes up.
Vidhu Vasudevan, Shashank Vivek Baratwaj
Built on Stripe / AgentPhone
DeepIndoors is an AI agent that helps users navigate indoor spaces such as airports, stadiums, and college campuses. Using LangGraph, the agent integrates with various software systems within an indoor space. For our airport MVP, this includes indoor schematics of the airport, tracking inbound and outbound calls, and information about loyalty programs, rewards, and terminal stores. These technologies can be integrated through the airport’s existing software systems or through our custom-built mapping and flight-tracking software (demonstrated in the demo video). The agent uses a hydration layer built with Moss.dev, enabling these integrations to propagate to the agent with sub-10 ms latency. Through AgentPhone, the agent is accessible via a telephone number, allowing everyone, including elderly, disabled, and international users, to access the system. Finally, with Supermemory, the agent’s indoor navigational abilities extend beyond a single call or even a single airport. The agent understands a user’s preferences across contexts, such as preferring halal food before boarding a plane or using AMEX cards, so that their experience is tailored to them at any airport. For example, when a user enters a new airport, the DeepIndoors agent remembers that they prefer halal food and can recommend nearby halal options where they can also receive 5% cashback using an AMEX card. Due to Supermemory’s cross-context memory across calls, DeepIndoors connects user behaviors between airports, stadiums, college campuses, and other indoor environments. As a result, DeepIndoors transforms indoor navigation into a personalized and dynamic experience for everyone, regardless of language, accommodations, or location.
Hanav Modasiya, Avik Chervirala, Jibril Suleiman, Ritvik Patra
Built on Moss / Supermemory / AgentPhone
Designing a room today takes 6 weeks: hours of Pinterest, a designer wait-list, multiple contractor quotes, and 5 separate invoices. Disi compresses that to a single phone call. User flow: the customer dials Disi's phone number (a real, working AgentPhone number). Disi answers, asks 3 short questions (room, vibe, budget), opens an iMessage channel for reference photos, then generates a multi-angle design proposal — moodboard, floor plan, three perspective renders, and an itemized shopping list — delivered as an iMessage carousel within 30 seconds. The customer reviews, says yes, and a Stripe Payment Link arrives. They pay via Apple Pay. From that moment, Disi autonomously pays vendors, schedules contractors, and assembles purchase carts across the web — all without further human intervention. Disi is the first product that needs both Stripe (humans paying agents) and x402 (agents paying anyone) — because she's the first agent that actually closes a real-world project, not a digital task.
Diana Munoz, Joel Timana
Built on Google DeepMind / Stripe / Browser Use / Moss / Supermemory / Sponge
We have no audio in my video demo. We just realized this. But I have the app deployed here: https://domainsdr.vercel.app/ DomainSDR is an autonomous domain broker agent. A domain owner enters a domain, price, floor price, and rules. The agent then works on its own: it researches likely buyers, finds contacts, sends low-volume personalized outreach using AgentMail, handles replies, negotiates within the seller’s limits, collects a Stripe deposit when a buyer is ready, and uses AgentPhone to schedule a post-deposit handoff call. The goal is to make domain sales less illiquid by turning each owned domain into an active outbound sales campaign, without requiring the owner to manually research buyers, write emails, follow up, or negotiate. It uses: - **AgentMail** for outbound and reply handling - **AgentPhone** for post-deposit calls and scheduling - **Stripe** for deposit collection - **Gemini** for buyer research, scoring, and reply understanding - **Apify / Browser Use** for buyer and contact discovery - **Supermemory** for long-term campaign memory and learned buyer/contact patterns - **Neon Postgres** for campaign state and transaction records In one sentence: **DomainSDR gives every domain owner an AI broker that keeps working until the domain gets a serious buyer, deposit, or handoff call.**
Carl Okpala, Jiyun Kim
Built on Google DeepMind / Stripe / Browser Use / AgentMail / AgentPhone
Fireside is the AI receptionist for the local service businesses SaaS has ignored: plumbers, dentists, vets, HVAC. They lose 30% of inbound calls to voicemail and can't afford a $3,000-a-month answering service. So we made it extremely simple. No engineers, no integrations, no agency, just a Google Maps URL. The owner pastes their URL. Fireside reads the business, provisions a real phone number, and configures an AI agent in under 30 seconds. The agent answers in the business's voice, books appointments, handles cancellations, and remembers callers across calls. We are killing agencies and serving unserved businesses.
Alexios Vasileiou, Konstantinos Ketselidis, Konstantinos Lampropoulos
Built on Google DeepMind / Browser Use / Moss / AgentPhone
Customer Support and Feature Requests are a thing of past
Het Patel
Built on AgentPhone
Fruit Market gives a physical fruit stand an AI operator that answers calls, recommends products, checks live camera inventory, takes payment, coordinates fulfillment, and restocks through human-approved supplier purchasing.
Amar Gandhi
Built on Google DeepMind / Stripe / Moss / AgentMail / Sponge
Job searching agent that receives user prompt and resume as input and returning top matching job postings.
Jake Kwon, Jerry Chen
Built on Google DeepMind / AWS / AgentPhone
GOFER is an autonomous errand operating system. Tell it what you need done, book a dentist, dispute a charge, order food and it dispatches a team of specialist agents that make real phone calls, browse authenticated sites with your profile, send emails, and prepare payments, all running in parallel. It stops before any irreversible action and asks for your approval. Every completed errand trains its memory so the next one is faster.
Carl Okpala, Jiyun Kim, Godson Ajodo, Elijah
Built on Google DeepMind / Browser Use / Moss / AgentMail / Supermemory / AgentPhone
GroundTruth is a founder decision engine for high-stakes operating questions. A founder can text in a problem, GroundTruth gathers company context, three CEO agents debate different plans, and the founder approves or rejects the recommendation before execution.
Anne Liu
Built on Supermemory / AgentPhone
ClaimPilot is an AI-powered claims assistant that helps users manage delayed or denied medical reimbursement claims through a simple iMessage-style conversation. Instead of forcing users to navigate confusing portals, paperwork, and follow-ups on their own, ClaimPilot lets them ask plain-English questions like “what’s missing?”, “status,” or “update claim.” Behind the scenes, ClaimPilot reads claim status, identifies missing documents such as EOBs or receipts, prepares claim updates, and sends proof back to the user once an action is completed. The system connects a user-facing iMessage agent with backend claim workflows, proof files, and status data so users can see what happened, what was done, and what needs to happen next. For the hackathon demo, ClaimPilot showcases a real phone-based flow where a user texts the assistant, receives claim status updates, triggers a claim update, and gets a proof link showing screenshots or documents confirming the update. The goal is to reduce claim friction, make reimbursement workflows easier to understand, and help users move from confusion to clear next steps.
Nikhita, Pavithra, Rohan
Built on Google DeepMind / Browser Use / AWS / AgentMail / Supermemory / AgentPhone
Haggle is an AI concierge that negotiates local services for you over the phone. Text Haggle what you need ("car detail in SF under $100 this weekend") and it handles everything: - Finds providers (Browser Use) - Calls them live (AgentPhone) - Negotiates pricing within your budget - Books, pays, and sends a follow-up once you confirm (AgentMail, Sponge) All via text/ a custom UI. How it works: 1. You text Haggle Send a request with your service, budget, timing, and preferences. 2. Haggle researches providers Using BrowserUse, Haggle scrapes and enriches local provider data, then ranks providers using historical interaction data from Supermemory. 3. Haggle calls providers Haggle places parallel outbound calls via AgentPhone. During the call, a voice agent runs a real-time negotiation loop: - Gemini generates one turn at a time - Waits for the provider’s response - Counters if pricing is over budget - Confirms price, timing, and payment method If nobody answers, Haggle falls back to cold email via AgentMail. 4. You review options Haggle presents providers ranked by: - Price - Availability - Reliability - Historical interactions 5. Booking + escrow You choose a provider and fund escrow using: - Card - USDC Powered by Sponge. The provider receives a payout link, and funds are released once you confirm the job is complete. The sticky loop: With Supermemory, Haggle remembers: - Preferred providers - Scheduling habits - Budget preferences - Communication style - Historical pricing across bookings It proactively suggests re-bookings when recurring services are overdue. The more you use Haggle, the less information you need to provide. Stack: - Next.js 15 on Vercel (serverless, ~120s max duration for long workflows) - Convex for real-time state (jobs, leads, calls, escrow, chat) - Gemini Flash for triage, intent parsing, negotiation summarization - AgentPhone for outbound voice calls + iMessage/SMS - AgentMail for email fallback + follow-ups - BrowserUse for provider scraping + enrichment - Supermemory for persistent user context (preferred providers, scheduling habits, price history, communication style) - Sponge for USDC escrow, card onramp, virtual card payouts, ACH via Bridge
Rishi Thakkar, Muzart Tuman, Evan Taylor
Built on Google DeepMind / Browser Use / AgentMail / Supermemory / Sponge / AgentPhone
HAGGL sends AI voice agents to negotiate procurement deals with suppliers in their native language — Twi, Hindi, whatever it takes — while you watch live. US companies lose billions in cross-border procurement because the supplier speaks Twi and the buyer speaks English. Deals die in translation. Brokers cost thousands. HAGGL fixes this: describe what you need, Gemini Deep Research finds real suppliers in Ghana and India, and HAGGL calls them simultaneously in their language. During the call, Claude Opus watches the transcript in real time and injects live market intelligence mid-conversation — Moss pricing data, historical supplier intel from Supermemory — so the agent always negotiates from a position of knowledge. When a deal closes, Sponge records the transaction, Browser Use fills the customs forms, and AgentMail notifies everyone. Every call makes the next one smarter. Claude Opus extracts negotiation patterns overnight, updates cultural playbooks, and promotes the best rebuttals into permanent memory. The agent that calls a Ghanaian textile supplier in month two knows everything from month one. We integrate all six sponsors: Gemini for the negotiation brain and supplier discovery, Moss for real-time market pricing mid-call, Supermemory for cross-call learning, AgentMail for autonomous post-deal emails, Sponge for transaction recording, and Browser Use for pre-call research and trade document completion. The phone book is open. HAGGL picked it up.
Aritro Bhattacharjee, David Nintang, Epaphras Akinola, Dhruva Vutukury
Built on Google DeepMind / Stripe / Browser Use / AWS / Moss / AgentMail / Supermemory / Sponge / AgentPhone
Hey Pookie! is a voice-first AI relationship concierge that helps couples turn missed hints into thoughtful real-world actions. Couples get an AI phone number they can call or text whenever they need help communicating, remembering, or planning something for their partner. Instead of acting like a generic relationship chatbot, Hey Pookie works like a practical “relationship ops” agent: it listens to a partner’s hint or request, understands the emotional context, remembers preferences, gently nudges the other partner, asks for approval, and helps complete the action. For example, if Maya calls and says, “My birthday is Friday, I’d love peonies and Thai food, but please don’t make it awkward,” Hey Pookie can extract the important details, remember that Maya likes peonies and hates lilies, check Alex’s availability, suggest a date plan, send Alex a gentle reminder, and get his approval before helping with flowers, dinner, payment, or an itinerary. Hey Pookie! turns relationship hints into handled actions — from “you should know” to “it’s already taken care of.”
Akshobhya Gupta
Built on Google DeepMind / Stripe / Browser Use / Moss / AgentMail / Supermemory / Sponge / AgentPhone
Hivemind automates the part of healthcare nobody wants to do: chasing insurance claims until they're actually paid. A billing specialist spends 30 to 45 minutes on a single claim. They log into payer portals, wait on hold with insurers, read off the same member and tax IDs every call, then chase the patient for whatever balance is left. A small clinic burns six figures a year on this work alone. Hivemind runs the whole loop on its own. A biller calls the agent and says "follow up on this claim." From there, Hivemind looks up how that specific payer behaves, then either calls the insurer and works the phone tree, or logs into the payer portal directly and clears 2FA by itself. Once it has the claim status and patient responsibility, it texts the patient a payment link to collect. The orchestration layer is what makes this work. An LLM orchestrator plans each claim, deciding per step whether to call, work the portal, or escalate, and delegates to specialized agents. Every run executes as a durable Temporal workflow, so a claim stuck on a 40-minute hold, interrupted by a crash, or hit with a retry resumes exactly where it left off. No replay, no lost state. The piece we're proudest of is restraint. When Hivemind hits a step only a licensed human can legally do, like a peer-to-peer clinical review, it stops, writes up the full context, and hands the task to the right teammate. It does not fake a result to keep the workflow moving. Stack: AgentPhone for voice and SMS, Moss for payer memory, Browser Use for portal automation, Stripe for payments, Temporal for durable workflows, LLM orchestrator on top.
Vedant Agarwal
Built on Google DeepMind / Stripe / Browser Use / AWS / Moss / AgentPhone
You tell the app where, when, and your budget. Our tool searches the web and contacts hosts and hotels to negotiate a discount on your behalf. It then ranks the best options — factoring in the negotiated discount — and shows you only the top picks. One click, and the agent completes the booking and payment for you.
Piotr Tyrakowski, Łukasz Kryczka, Michał Zajączkowski, Wojtek Krupiński
Built on Browser Use / AgentMail / Supermemory / Sponge / AgentPhone
AI agents that run property management... in a pixelgame.
Jakub Sobolewski, Piotr Gatkowski
Built on Google DeepMind / Stripe / Browser Use / AWS / AgentMail / Supermemory / AgentPhone
David is your personal accountability coach, calls & checks in with you every day, manages your tasks on Notion, send money to organization you hate if you follow hit goals, books things for you that you've been putting off using the budget you give him
Ahmet Ilten
Built on Google DeepMind / Browser Use / Sponge / AgentPhone
Jenny is an AI Marketing Officer. She finds the best fit between a brand and influencers. It does market research based on just the product name and a short description (from a call with AgentPhone), and uses that information to find the best content creators that can promote the product. Once we have the influencers, we send them all emails and follow up with any questions they have and automatically close the deal. As a manual process this could take weeks, and requires niche specific knowledge, now the brand only needs to say yes and pay a Stripe invoice. Jenny handles the negotiation and content ideation.
David Gregory, Homit Dalia
Built on Google DeepMind / Stripe / Browser Use / Moss / AgentMail / Supermemory / AgentPhone
It is a job application agent using Moss. Moss's semantic search enables faster resume and job posting matching percentage.
Jake Kwon, Jerry Chen
Built on Google DeepMind / AWS / Moss / AgentPhone
agent that can actually
Maggie Chen
Built on Browser Use / AgentMail / Supermemory / AgentPhone
Live demo: https://threat-vector.vercel.app/ Kairos is a real-time AI threat intelligence platform for school district tip lines. A student calls or texts a dedicated school number. Kairos answers 24/7, transcribes the call in 70+ languages via Gemini Live, and runs three independent AI models in parallel: a primary semantic classifier extracting 22 structured fields from the transcript, an independent second-opinion model for consensus verification, and a Bayesian Monte Carlo engine running 500 probability simulations to produce a calibrated threat score with a 95% confidence interval. Within 12 seconds of call completion, the school principal receives an SMS alert and the district safety officer receives a structured HTML intelligence brief. Every tip is geocoded onto a live GPS heatmap, named subjects trigger a multi-source background check across DuckDuckGo, CourtListener, and Bing, and every autonomous agent action is logged as a Sponge micropayment transaction. The entire pipeline runs autonomously with no human in the loop, from inbound call to principal notification. Built on AgentPhone, Gemini Live, Anthropic, Gemini Flash, Supermemory, Moss, Sponge, AgentMail, and Supabase Realtime.
Ishaan Samantray
Built on Google DeepMind / Browser Use / AWS / Moss / AgentMail / Supermemory / Sponge / AgentPhone
Permit AI for SMB
Sukanya Moorthy
Built on Google DeepMind / Browser Use / AgentPhone
Scientific procurement is slow, fragmented, and very manual, forcing researchers to spend valuable time searching for suppliers, comparing products, emailing vendors, and tracking procurement workflows. Laab handles all of this, resulting in faster sourcing, reduced operational overhead, and a centralized system that helps scientists focus on what actually matters. Laab intakes AgentMail/AgentPhone transcripts, turns them into a Supermemory vendor dependency graph, and uses this to answer any questions a scientist might have through the AgentPhone number. Can't find the number? Browser Use scrapes the contact information of the vendor for itself. The agent also uses sponge/Stripe integrations to complete transactions autonomously (or not, given the proper permissions).
Vasilisa Baginskaya
Built on Google DeepMind / Stripe / Browser Use / AWS / AgentMail / Supermemory / Sponge / AgentPhone
Landline is a voice-based middleware platform for the hospitality and property management space. Tenants can call a voice agent to report issues, and the system handles the full interaction by collecting details, logging the request, and routing it to the right service worker. After each call, tasks are automatically created and tracked on a central platform. A key part of this is how work gets assigned. Our agent automatically selects the right service worker for each job based on the nature of the issue, then notifies them instantly via AgentMail. Service workers can see all their assigned tasks directly on the dashboard, keeping everyone aligned without any manual coordination. On the landlord side, we built a dashboard to view all tickets, track progress, and see analytics on recurring issues and overall property performance. A big part of what makes Landline powerful is how we used Supermemory. The voice agent doesn't just handle one-off calls it actually learns over time. It remembers who the tenant is, what unit they live in, and whether an issue has come up before. This lets it recognize patterns, handle repeat problems more efficiently, and make interactions feel much more personalized. Additionally, we used Supermemory on the landlord dashboard to dynamically tailor insights based on what the landlord asks, instead of showing static analytics. We integrated AgentPhone for both inbound and outbound communication, optimized to also call tenants with updates or upcoming appointment reminders. Overall, Landline makes property management workflows faster, more automated, and more personalized for both tenants and landlords.
Sriram Manikandan, Rishit Agrawal, Aman Bollam
Built on AgentMail / Supermemory / AgentPhone
AI Voice Agent that sits on your website 24/7 to call and bring back all the abandon checkout. It gets implemented with one line of script tag.
Sohazur Islam, Shanzila Ahmed, Maryam Alabbar
Built on Moss / Supermemory
hire an ai receptionist over the phone
adele beitvashahi
Built on Stripe / Browser Use / AgentMail / AgentPhone
Leakly is an autonomous property maintenance dispatcher that handles the entire repair lifecycle without human intervention. When a tenant reports an issue via chat or email, AI classifies the problem, selects and contacts the best vendor, schedules the appointment based on the tenant's availability, and pays the vendor in USDC on-chain via Sponge. Every step — from intake to payment receipt — is automated, with the tenant and property manager able to track progress in real-time.
Arjun Chidambaram, Dante Kakhadze
Built on Google DeepMind / Browser Use / AgentMail / Supermemory / Sponge
AI operating system for personal injury law firms where AgentPhone agents work the phones in parallel, Moss does sub-second semantic lookups mid-call, Supermemory remembers every client and case the firm has ever touched, Browser Use scrapes fresh precedent off court sites live, AgentMail fires the intake follow-up before the client hangs up, and AWS holds it all together, like Better Call Saul if Saul never slept and never missed a call.
Arman Luthra, Fardin Hoque (Kazi Fardinul Hoque)
Built on Browser Use / AWS / Moss / AgentMail / Supermemory / AgentPhone
Your entire AI marketing team in a phone call.
Chandra Sravanthi Kaki
Built on Google DeepMind / Browser Use / AgentMail / Supermemory / AgentPhone
Today we say goodbye to hold music, ai chat support bots, and get access to real humans now! MeWanthuman is a cross-platform tool that connects you directly to a real person at customer support. It races two channels at once, a phone call navigating the automatic phone menus and a cloud browser-use agent finding live chat and reaching a human. whoever reaches a human first, wins. Works everywhere: iOS (embeds into any of your apps, no shortcuts , uses dynamically loaded library easy to inject), Chrome (extension), Safari (iOS extension), a Mobile App, and a Website. just enter any phone number and it takes over. Every successful connection gets indexed into memory via Supermemory, so the next time you call the same company, the agent already knows the fastest path to a human, which makes it avoid trying dead ends.
Sina Matian
Built on Browser Use / Supermemory
Every year 6 million students try to study abroad and most fail , not because they can't get in, but because the application process is a nightmare and real agents charge thousands upfront. Agera is an AI agent that handles the entire application process over a phone call, so even a student in rural Kenya with no internet can access the same help and be able to apply with ease . We charge nothing upfront — we collect $2,500–$7,000 from the university only when a student successfully enrolls.
Nathaniel Deryl Wankio
Built on Google DeepMind / Browser Use / AWS / AgentMail / AgentPhone
Miranda is the first voice-native legal marketplace. Today, 350M Americans have no number to call when they need a lawyer, and law firms lose 40% of their inbound leads to voicemail every missed call is a $50K case walking next door. Miranda answers the phone 24/7. In under a minute on a single call, she holds a real conversation grounded in actual California statute (no hallucinations — KB retrieval mid-call via Supermemory), qualifies the case, scores it 0-10 with reasoning and red flags, drafts a personalized iMessage and email follow-up, and routes the scored case to the attorney's dashboard. One click "Take Case" auto-notifies the client across SMS and email. She remembers every caller by name across calls. We don't generate leads. We deliver signed clients. Try it live: call +1 320 684 6220 from any phone, anytime — Mira will pick up. Built on AgentPhone (voice + iMessage), AgentMail (email + threading), Supermemory (memory + KB), and OpenAI (reasoning). Multi-tenant Next.js admin, FastAPI backend, Postgres. Miranda placed 1st runner-up at the Stanford CodeX LLM × Law Hackathon #6, and was invited to join Antler off the back of it. Demo: https://www.loom.com/share/665942b0e9424de6906cadd405a9cbc6 Code: https://github.com/kollaikal-rupesh/Miranda
Kollaikal Rupesh
Built on AgentMail / Supermemory / AgentPhone
Measure your work and life style and give event suggestions that maximize your work life balance
Maryna Iseianko
Built on Google DeepMind / AgentMail / Supermemory / Sponge / AgentPhone
forgeOS is a workspace for the agents that run your company. Engineering, marketing, research, social, support — each one a teammate, all sharing context, all working in parallel.
Jaeyun Ha
Built on AgentMail / AgentPhone
NextAction is a unified customer action engine. When a Stripe invoice.payment_failed event fires, it runs a real-time decision pipeline: hydrates customer context, routes to an A/B experiment arm, scores action candidates, and executes — either an AgentPhone voice call + SMS payment link for customers with phone consent, or an AgentMail email for those without. The call is real, the Stripe webhook is real, the recovery is real. Every decision is fully traceable from signal to outcome, and experiment results surface as one-click approval proposals to update future policy. URL: https://nextaction-git-main-shingo8843s-projects.vercel.app/
Shingo Nagano
Built on Stripe / AgentMail / AgentPhone
Nexus is the Twilio for AI agents. Give it a mission in plain English — "negotiate a group hotel rate," "send cold outreach to 20 prospects," "cancel my subscription and get a refund" — and it autonomously researches contacts, picks the optimal channel (phone, email, or chat), handles the entire conversation, follows up, and reports results. Under the hood: dedicated microservices for each channel, a custom chat engine that speaks native chatbot protocols (no browser automation), parallel execution across channels, persistent memory across sessions, and an API so other AI agents can use it as their communication layer. Built with AgentPhone, AgentMail, Moss, SuperMemory, Google DeepMind Gemini 3.1 Pro, and deployed on AWS. Pick whichever length the form needs. Want me to trim it further?
Sreekanth Putta
Built on Google DeepMind / AWS / Moss / AgentMail / Supermemory / AgentPhone
First-line defense for production on-call: the AI picks up the page, triages the incident, runs the runbook, and resolves it. Humans only get woken up when the agent can't.
Banashankar Veerad
Built on AgentMail / AgentPhone
The global conversational AI market is projected to surpass $41 billion by 2030 (Grand View Research, 2024), driven by increasing demand for AI-powered automation, and real-time business operations. At the same time, customer expectations are rapidly changing — 90% of customers say an immediate response is important when contacting a business (Help Scout, 2024). As Tom Blomfield recently described in YC’s Company Brain thesis, “The biggest blocker to AI automation of companies is no longer the models… now the blocker is the domain knowledge.” Modern businesses operate with fragmented information spread across calls, emails, support tickets, CRMs, and team conversations, making it difficult for AI systems to reliably execute real operational tasks. Opelo is an AI-powered customer operations agent built to become part of this emerging “company brain” layer for businesses, helping companies automate conversations, calls, bookings, and payments within one intelligent workflow. Instead of relying on fragmented tools across messaging, scheduling, and customer support, businesses can use Opelo to automate customer interactions end-to-end from answering inquiries and handling voice calls to confirming bookings and processing payments in real time. For the current demo flow, customers can interact with an Opelo agent through text or voice, complete a booking process, and make small cent-based payments directly during the interaction, demonstrating how AI agents can move beyond passive chatbots into fully actionable business operators
Raina Zab, Precious Inyang
Built on Google DeepMind / Stripe / AgentMail / Supermemory / Sponge / AgentPhone
Panacea is an agentic AI persona that captures context for therapists and their clients to provide 24/7 support when the therapist is unavailable. We want to solve the problem of mental health specialists being so hard to get.
Zuyi Zhao, Dhruvil Bhatt
Built on Browser Use / AWS / Moss / AgentMail / Supermemory / AgentPhone
Med-legal
Caleb & Luke Winningham
Built on Browser Use / Moss / AgentMail / Supermemory / AgentPhone
Peripheral is an agent-native HUD for smart glasses. Peripheral gives AI agents a visual language: glanceable status icons, approval cards, meeting briefs, live task updates, and contextual widgets rendered directly onto monochrome waveguide displays. Agents feel present in the glasses, can update the HUD as they work, and can ask for approval at the exact moment the user needs to decide.
Karim Yahia
Built on Google DeepMind / Stripe / Browser Use / Moss / AgentMail / Supermemory / Sponge / AgentPhone
Pigeon is a support and diagnostics platform that helps EV charger operators improve reliability, reduce downtime, and cut repair costs. When a driver reports a broken charger, Pigeon captures and transcribes the call in real time, combines it with live charger telemetry, usage history, and component data, then diagnoses the likely issue. If the problem can be fixed remotely, Pigeon helps resolve it immediately; if not, it automatically generates a detailed work order and sends it to the right technician with the context they need. By turning vague customer complaints into actionable diagnostics, Pigeon helps operators avoid repeat visits, save money, and deliver a charging experience EV drivers can trust.
Senara Millawabandara, Suhani Shokeen, Atin Singh, Sina Vaghefi
Built on Google DeepMind / Moss / AgentMail / Supermemory / AgentPhone
Parody - Pitch YC - answer is always the same.
Spb176
Built on AgentPhone
Multilingual insurance agent
Anna McGovern, Bek Hamit
Built on Google DeepMind / Browser Use / AgentMail / Supermemory / AgentPhone
Poxi is Clicky but with action, It acts on you're behalf on your computer read your email summerize and you can interacti with it through text
Yeabsira Teshome
Built on Google DeepMind / Stripe / Browser Use / AWS / AgentMail / Supermemory / AgentPhone
ProxyCall - the AI that makes the calls you're too scared, too busy, or too cheap to make yourself One-line pitch. You type the intent. ProxyCall picks up the phone, has the conversation, ends it gracefully, and hands you a full debrief in one click. The calls every human is avoiding right now "Hey - I think we met at the coffee shop last Tuesday. Would you want to grab dinner Friday?" - and somehow your assistant is doing the asking, with the exact tone you specified, while you stare at the ceiling. "United just mashed up my connection in Newark. I'd like a full refund and miles for the trouble - and I'm not hanging up until we get there." "Hi, I'm looking at the Marriott downtown for two nights. Your rate online is $389. I'm going to need that closer to $260 or I'm walking to the Hilton." "Table for four, Saturday, 8pm, somewhere on the patio if possible. If they're full, ask what Tuesday looks like." "My landlord still hasn't fixed the AC. Call him, be polite, but make it clear we're documenting this." "Tell my mom I love her and I'll be home for Thanksgiving. She'll cry. Be patient." You write it in one sentence. ProxyCall dials. It says exactly what an honest, human-paced assistant would say - never pretending to be you, never lying, never bulldozing the person on the other end. Then it hangs up at the right moment and drops the entire conversation in your dashboard: Transcript. Every line, in order, with timestamps. Outcome. goal_met / partial / follow_up_needed / declined / opted_out. Emotional read. Was she warm? Was the airline agent annoyed? Was the hotel desk bluffing? Action items. "Send your ID to the desk before 6pm." "Call back Tuesday." "Bring flowers." A draft message you could send the recipient right now - already in your voice. One tap to copy. That's the whole loop. Intent in. Call placed. Story back. One click. Why this wins Call My Agent Every other team at this hackathon is shipping the same agent: a cold caller that sells. The cold-caller market is a knife fight with twenty other YC companies. We took the inverse — and we think it's bigger. These are calls a normal human would pay $5 each to skip, every week, for the rest of their life. That's not a B2B SaaS pipeline. That's a consumer product the entire room would install before they got back to their hotel tonight. What's actually built, not pitched A working Next.js 15 app with two demo paths a judge can drive themselves: Real telephony. AgentPhone + Gemini 2.5 + ngrok. Put your number in /calls/new, write the brief, hit submit. Your phone rings. You talk to your own agent. Zero-account simulator. From the call detail page, drive the recipient side and watch the agent reply through the exact same /api/agentphone/webhook handler production uses. The guardrails are identical. The demo works even if the WiFi dies. Three engineering choices a judge will respect Honesty is a system property. The model returns strict JSON every turn - { reply, shouldEndCall, optOut, askedToBookCall, revealedAi }. If a recipient asks "Are you a person?" it cannot lie; the prompt makes that impossible, and the contract makes it auditable. Latency is treated like a deadline. AgentPhone's webhook window is unforgiving. Gemini 2.5's "thinking" mode blows it. We ship a separate voice path - trimmed system prompt, clipped owner context, 128-token reply cap, thinkingBudget: 0, and a 22-second hard wall-clock race with graceful-fallback lines. The callee never hears a hung socket. Consent is permanent. When the hotel rep says "please stop calling me" — or your ex says it, more painfully - the model emits optOut: true, we write the number to data/opt-outs.json, and the next attempt to call that number 409s before AgentPhone is ever contacted. The consent ledger survives restarts. It is the whole moral and legal moat of consumer voice agents, and it works today. The sponsor surface — this thing is built for the ecosystem in the room AgentPhone — every outbound call, every webhook event. Google DeepMind / Gemini 2.5 — live voice (Flash-Lite) and post-call summaries (Flash). Supermemory — drop-in for cross-call recipient memory. The schema for relationships already exists in UserProfile. AgentMail — the "suggested follow-up message" is a one-line swap away from an actual email send. Stripe / Sponge — the Doer track is "book and pay." We already capture the action item; wiring payment is a single endpoint. Browser Use — when the airline rep says "please go to united.com/refunds and submit a form," the agent should not hang up. It should hand off to a browser agent and finish the job. That's the multi-channel "fixer" play, and our architecture has clean seams for it. Why this team ships The product is small, legible, and ruthless about contracts: lib/prompts.ts - the only file that talks to the model. lib/agentphone.ts - the only file that talks to telephony. lib/storage.ts - a 10-function async surface; swap to Postgres in an afternoon. Three seams, three contracts, no hidden glue. That is what shipping looks like. You write the intent. ProxyCall makes the call. One click and the whole story comes back. Ask her out. Get your refund. Lowball the hotel. Book the table. Call your mom. Pick up the phone. Links for hackathon(its in continuation) Link 1: https://www.loom.com/share/7e3515e2ed334f749493cf0e7a36b14f Link 2 : https://www.loom.com/share/525b959c0b0f4d5bb207aebb949cbfe4
Nirav Rohra
Built on Google DeepMind / AgentPhone
A patient who can't describe their pain doesn't get treatment for it. For the millions calling US hospitals in their second language, the choice is wait twenty minutes for an interpreter, hand the phone to your child, or hang up. ReadBack picks up immediately, in any language, and gives every patient the patience to say what's actually wrong — then hands the full conversation to a qualified human the moment the call touches anything clinical, so the patient never has to start over. Patience for patients. Care, in every language.
Kuanze Ma, Alison (McCarthy) Cossette, Jonathan [email protected]
Built on Google DeepMind / Stripe / AgentMail / Supermemory / Sponge / AgentPhone
Recover money for consumers (flight / hotel / order delay & cancellation and many more)
Joe He
Built on AgentPhone
Robin — Every hero has a sidekick. You need one too. You phone Robin. Say five words: "Cancel my gym membership." Robin runs a real discovery dialogue (brainstorm → plan → confirm) until the goal is unambiguous, then pulls the actual cancellation law off the live web while you're still on the call. It dials the gym, t-mobile, amazon, everything you nee to contact. Act on your behalf and work the retention line with a negotiation playbook: like Chriss Voss (ex FBI negotiator) would. Then call you back or followup with you via messages/email with the actual status! Boom! 2hrs of your life back!
Francesco Rosciano
Built on Browser Use / Moss / Supermemory / AgentPhone
Find, order, and pay for food via text
Joe LaChance
Built on Stripe / Browser Use / AgentMail / Supermemory / Sponge / AgentPhone
What Siri should have been. AI Assistant that lives on your Mac, helping you learn new things and handle mundane tasks that aren't worth your time.
Dhiva Krishna
Built on Google DeepMind / AgentMail / AgentPhone
Scall takes revenge on scammers. With over 30 billion scam calls and 2.5 trillion scam emails sent every year, scammers cost Americans' time, drain phone data, and steal billions of dollars. Scall fights back. Simply forward suspicious emails to Scall, where our AI voice persona, "Marge," will call and engage scammers in long, believable conversations. Every minute a scammer spends talking to Scall is time they can’t spend targeting real victims---like your sweet old grandmother. By wasting scammers’ time before they can extort money, Scall protects users while saving both valuable time and hard-earned money.
Nico Zametto
Built on Google DeepMind / AgentMail / AgentPhone
The average car dealer misses ~10 calls a day. A brake pads replacement job is $400 — that's $4,000 lost before lunch. ServiceLine answers every call, schedules the appointment, invoices, and collects payment. Zero humans. It also proactively calls customers due for upcoming service and finds new leads from Reddit. Built with AgentPhone, AgentMail, Browser Use, and Stripe.
Abdalla Odat
Built on Google DeepMind / Stripe / Browser Use / AWS / AgentMail / AgentPhone
An AI phone-screening agent that uses AgentPhone to automatically call and screen applicants, generating a structured candidate deliverable for your ATS in real-time and leveraging Supermemory to query resumes and prior context so each interview is tailored to the applicant.
Ayush Kumar
Built on Supermemory
Computer-use agent fleets for enterprise workflows. Legacy systems have no APIs, and browser agents stop at the tab. Solitude runs computer-use agents in the cloud that operate any application: native, browser, or desktop app. Like a new hire you trained once, you record a workflow in 15 minutes, and the agent runs it 24/7 on our infrastructure, with credentials and 2FA handled. Solitude will enable small enterprises to bypass lengthy & expensive API integrations, and large enterprises to execute repetitive and time-consuming manual tasks over a digital interface.
Adrian Coutsoftides, Yousef Kassem, Matthieu Thienpont, Stefan Abrudan
Built on Google DeepMind / Browser Use / AgentMail
It costs $51.5 billion just to fix the backlog of repairs sitting in America's rental housing stock right now (Federal Reserve Bank of Philadelphia). And 39% of property managers say they spend 20+ hours a month just playing phone tag to get those fixes done (Buildium 2024). A pipe bursts at 2am—the tenant calls, the manager scrambles to reach contractors, nobody picks up, and two days later the ceiling is ruined. Handle fixes this end to end. The tenant calls in the issue—BrowserUse pulls nearby contractor information, and AgentPhone dials them in parallel, with Moss and SuperMemory feeding the agent context on past jobs and contractor history to negotiate the best rate. If the tenant can't describe the problem clearly, the agent requests a video and Gemini analyzes it to diagnose the issue and hand that context back to the agents. If no one picks up, AgentEmail automatically reaches out to the next available contractor. When the job is done, Sponge auto-pays the contractor from the landlord's account via Stripe—no invoice chasing, no back and forth. The entire process is done without the property manager having to pick up their phone.
Alex Battikha, Rishab Motgi, Nicolas Dos Santos
Built on Google DeepMind / Stripe / Browser Use / Moss / AgentMail / Supermemory / Sponge / AgentPhone
Riley is a single phone number that both parents and kids text. Parents check grades, monitor browsing, manage calendars, and approve spending. Kids request money and relay messages. Everything happens over iMessage ~ no app, no login, no screen. Just text.
Jacob Fu, Sebastian Jurado, Gabriel Dennis Shores
Built on Browser Use / Supermemory / Sponge / AgentPhone
The Dream Machine leverages your personal knowledge graph (your worldmodel powered by supermemory) to track your dreams, explain why you had them and provide you with a deeper understanding of your own subconscious. Dreams submitted via iMessage (AgentPhone) are enriched with relevant metadata then ingested into your worldmodel where semantic fields connects ideas and concepts. Those ideas and concepts, born of the influences in your life, are then connected to your dream via graph relatedness. The Dream Machine then synthesizes a response, texted and printed on a thermal label, for you to document your dream and explain to you why you might have had the dream you had last night. As people and agents become more intertwined, everyone will have rich, personal knowledge graphs that hold their personal context - the dream machine will not only help you track your dreams, but it will tell you something about yourself you may not have yet realized.
Alex Funk
Built on Google DeepMind / Supermemory / AgentPhone
Meet Sophia: The AI recruiter that handles all your candidate pre-screening for your company.
Jose Mendoza
Built on Google DeepMind / AgentPhone
Traction is autonomous customer development for tech founders. You connect your GitHub repo and product docs, and Traction handles the entire pipeline of finding, contacting, and booking beta tester interviews — completely autonomously. You wake up to a calendar full of booked user interviews without writing a single cold email or scheduling a single call. The unfair advantage: Browser Use connects to the founder's real, logged-in Chrome session (7,400+ synced cookies across LinkedIn, GitHub, Reddit, X). Anti-bot systems can't flag it because it IS the founder's authenticated session — making research possible at a scale and depth that no scraping-based outbound tool can match. How it works end-to-end: Claude Opus ingests the founder's GitHub and generates the ICP, value propositions, and pitch angles. Browser Use researches prospects in parallel using the founder's authenticated browser sessions, extracting structured data from their GitHub commits, LinkedIn posts, and Reddit discussions. AgentMail sends genuinely personalized outreach (referencing specific projects and commits — not "{{name}}" templates) with a Calendly link plus an option to reply with available times. A polling loop checks AgentMail every 2 seconds for replies, runs them through Claude to detect intent and extract dates/phone numbers, and then either auto-books on the founder's real Google Calendar with a Google Meet link, or triggers an AgentPhone outbound call when the prospect says "call me." On the live call, the agent has full prospect context loaded from Supermemory at ~50ms latency and can answer product questions in real time using Moss for sub-10ms search over the founder's product docs. After every interaction — email reply, voice call, meeting — the transcript and extracted insights flow back into Supermemory per-prospect containers and a shared swarm-insights container so the system gets smarter with every conversation. Every sponsor is used in a load-bearing way, not as a checkbox: AgentPhone (voice swarm with greeting, conversational AI, and outbound call triggering on email reply); Browser Use (authenticated Cloud sessions with synced founder profile, structured Pydantic extraction with text-parse fallback); AgentMail (real email send/receive, webhook + 2s polling fallback, dedup, auto-prospect creation for unknown senders); Supermemory (per-prospect containers with entityContext-guided extraction, shared swarm-insights container, profile retrieval, conversation ingestion after every touchpoint); Moss (per-prospect indexes plus a shared product-knowledge index, hybrid alpha tuning, sub-10ms voice-call queries); Sponge (multi-chain agent wallet with x402 paid_fetch for autonomous research API payments); Stripe (usage-based billing meters for interviews booked and outreach sent). Plus Google Calendar OAuth for real freeBusy queries and Meet event creation, and Claude Opus 4 as the orchestrator brain that ties it all together. Built in one day at the Call My Agent Hackathon at YC. 8,500 lines of code across a Python FastAPI backend and a Next.js dashboard. Verified end-to-end with real API calls: real personalized email sent, real Google Calendar meeting booked (Google Meet link generated), real outbound phone call triggered from email reply intent, real Browser Use session extracting GitHub data with the founder's profile. The demo shows a prospect replying "book Wednesday 3pm" and the system auto-booking on the calendar within seconds, then replying "call me at this number" and the prospect's phone ringing within seconds — fully autonomous.
Elijah
Built on Google DeepMind / Stripe / Browser Use / AWS / Moss / AgentMail / Supermemory / Sponge / AgentPhone
Virtual host for restaurants that answers the phone and takes reservations
Stepan Arsentjev
Built on Stripe / Moss
You call or text or email the agent a task. It quotes a price and sends you a payment link. As soon as you pay it, it gets done.
Reza Jamei, Maria Reiling, Lara Hoyem
Built on Google DeepMind / Stripe / Browser Use / AWS / AgentMail / AgentPhone
An agentic scheduler - u just name the goal you want to achieve + duration. Find open slots in your calendar, find resources, interact with real world and fill up your calendar with things you need to do in order to achieve it. We provide each goal as an agent that runs in background with all tools like agent mail, phone number, browser use and it interacts with real world to propose necessary todo in order to make you achieve the goals
Syed Mustafa
Built on Browser Use / AgentMail / AgentPhone
Zyra is an AI-powered healthcare coordination assistant that helps users discover the right specialists based on symptoms, location, insurance, and urgency. It combines voice intake, AI-driven care matching, live web research, and automated appointment coordination to simplify the healthcare navigation experience.
Anjali Gupta, Shikhar Brajesh
Built on Google DeepMind / Browser Use / AgentPhone
295 shots from kickoff through the closing ceremony, captured by Rome Media Agency.
Thank you to the partners who made this possible.
