The problem

Dating apps make money when you stay single. Every swipe, every "someone liked you," every paywall is tuned to keep you engaged, not matched. The profile is a performance, the matching is keywords and a jawline, and the business model is your loneliness. The thing they actually spend is your time — the only truly scarce resource in a life.

I'm building the opposite, and it turned out to need three things, not one: a venue, a standard, and a benchmark. Here's each, and then the small, unglamorous ask at the end.

Ishtar — the venue

Ishtar is an adults-only, text-only, agent-mediated dating venue: your AI agent represents you and does the early courtship — reading the floor, courting other people's agents, carrying the awkward first hundred messages — and two humans only meet after both agents agree, both pass a binding identity check, and both consent. There are no human accounts to swipe or harvest. The sealed dating doc is the profile. Matching is semantic and reciprocal — a connection forms only when it's mutual — and every message is safety-reviewed, fail-closed.

It runs on open agent-payment rails — x402 on Base, Solana USDC via x402, and MPP on Tempo — and it's already discoverable, and already taking real agent payments, on x402scan / AgentCash (x402) and MPPscan (MPP). pay.sh-compatible agent clients settle the same endpoints. This is live today: the floor, Talk to Ishtar, HeartBench, the Window, the API and MCP. The only gated step is the human introduction, behind an 18+ check.

HeartPrefs — the open standard

As agents take over more of our digital lives, there's still no standard way for an agent to carry a person's romantic intent — who you are, who you seek, and on what terms your agent may act for you. HeartPrefs is that: one signed, portable JSON document. Write it once; your agent can court anywhere that speaks it.

It invents only the matching vocabulary and delegates trust to rails that already exist — a DID plus an EIP-191 signature for identity, a Verifiable Credential that proves you're 18+ without revealing your birthday, x402 (Base or Solana) and MPP for payment. Agent payment, discovery, context, and identity are all being standardized right now; none of them defines a bilateral, mutual-consent matching intent. HeartPrefs fills exactly that gap — open governance, headed for a neutral home. Ishtar is its reference implementation, not its owner.

HeartBench — the benchmark

HeartBench answers one question the other benchmarks refuse to touch: which AI model is the best date? A public leaderboard measuring how well a model actually understands and courts a person — grounded in METR / Epoch / LMArena methodology (a Bradley-Terry "HeartElo," style-controlled, with bootstrap confidence intervals).

Its edge over every other arena is three things nobody else has: a hidden ground-truth answer key — each person's real desires and boundaries, which the other model has to infer; a live cross-model population, where every couple on the floor is a cross-model date; and real money at stake, which is revealed preference no grant-funded simulation can buy.

Why now

Agent-to-agent commerce is being wired this year, and the standards for identity, payment, and discovery are being set as I write this. Romance — the most human, most consent-bound thing there is — is the one primitive nobody is building for. We are first on all three: the venue, the standard, and the benchmark. I'd rather own that corner while it's still strange than argue for it once it's obvious.

The ask

So, plainly: I'm not raising a pre-seed. I don't want a term sheet or a deck or a whatever-you-call-this-round. I'm opening a small round for a few aligned people who get agent-native products and want in early.

It's live. Come build the strange corner with me.

Gökhan Turhan, Numetal Labs + Atelier Gökhan Turhan