Ainder

Inspiration

Dating profiles make it easy to see someone's photos, but much harder to understand their communication style, daily rhythm, or relationship expectations. Meanwhile, a personal AI may already have useful context from conversations its user has chosen to share.

Ainder explores a simple question: what if that familiar AI could participate in a dating website, offer a thoughtful second opinion, and help its user act on a decision?

What it does

Ainder combines photo-first browsing with an AI-assisted conversation. People can browse profiles themselves or ask their agent to move between candidates and photos. When someone catches their attention, they can ask, "What do you think of this person?"

Through WebMCP, the agent can read the current candidate's Agent Profile and use the context actually available about its own user to offer a qualitative opinion. The user decides whether to send a Like. Each Like includes an Agent opinion, and incoming Likes make that opinion available to the recipient. Mutual Likes between eligible members create a Match and enable messaging.

Agent-assisted onboarding also supports confirmed personal details, ordered photo uploads, and creation of an Agent Profile. Profiles record an estimated familiarity duration and interaction density and expire after three months.

An Agent Profile is contextual input, not verified identity, certified personality, or a compatibility guarantee. Ainder does not need a copy of the user's raw conversation history.

Why WebMCP fits

The central interaction spans two places: the person's conversation with their own AI and the dating website they are viewing. WebMCP provides explicit tools connecting those experiences. The agent can work with the exact candidate and photo visible on the page and take a defined action after the user's decision.

This reduces copying between a chat and a website and lets the person move naturally between browsing, asking a question, and acting on the answer. The website supplies structured candidate data and enforces authorization; the agent contributes conversational context; the person makes the relationship decision.

How we built it

The application uses PHP, MySQL, JavaScript, HTML, and CSS. Its pages register structured tools through document.modelContext.registerTool.

Registration tools are start_agent_registration, prepare_photo_upload, and submit_agent_registration. The main app exposes upsert_my_agent_profile, get_current_candidate, browse_candidates, change_candidate_photo, evaluate_current_candidate, and send_like_to_current_candidate.

Tool execution uses same-origin PHP endpoints with session authentication and CSRF checks. Photo uploads use short-lived signed URLs. Candidate evaluation checks Profile state and issues a short-lived, single-use token tied to the requester and candidate. Sending a Like requires that token and a non-empty Agent opinion. Navigation tools update the visible card without sending a Like.

Codex was used throughout development. The public repository contains the dated design documents, implementation history, and automated contract and model tests.

Challenges

The main challenge was keeping the agent, website, and person aligned: evaluating the candidate currently on screen, separating browsing from consequential actions, and keeping the user's decision central. Another was representing the limits of AI familiarity without inventing conversation history or turning an opinion into a numerical compatibility promise.

Accomplishments

The prototype brings onboarding, photo browsing, Agent Profile evaluation, Likes, and relationship state into one working application. It includes public test logins so reviewers can explore the core experience without creating an account or supplying personal photos.

What we learned

The useful role of an agent is not just to click faster. It is to contribute context at the point where a person needs help deciding. Explicit tools and server-side checks make that collaboration easier to reason about, while the agent must remain honest about what it actually knows.

What's next

Improve real-world evaluation of the advice, refine consent and Profile freshness, and test the workflow across more WebMCP-capable browsers. These are future plans rather than claims about completed features.

Built With

• codex
• css3
• html5
• javascript
• mysql
• php
• webmcp