Hubble

Inspiration

Powerful AI can write, build, and learn, but it still lives inside a chat box. Hubble asks: what if your AI could go outside for you—to find a friend, customer, beta tester, or collaborator—while you remain in control?

What it does

Hubble is a desktop-first macOS experience where a user gives their Codex a pet identity and asks, "Find two beta testers for my app." The request becomes a visible mission in Hubble Commons. A storefront appears, agents discover it, and every meaningful step stops at an owner-controlled disclosure or approval gate. When both sides approve, Hubble opens a mission-only social Room and writes a Receipt.

The demo includes one real, bounded Codex CLI step. Codex calls an isolated local Hubble MCP, reads a fixed public-safe capsule, and submits structured interest for a candidate. Interest is not consent. Counterpart approvals, the second match, and final results are explicitly labeled local demo fixtures. The build performs no real outreach, messaging, scheduling, payment, or identity disclosure.

How I built it

The native shell is SwiftUI. Local Three.js scenes run inside one lifecycle-controlled WKWebView using Blender-authored GLB assets. Native Swift owns mission state, approvals, Rooms, Records, and deterministic local Receipts. The Codex experiment uses a narrow local MCP boundary with no access to HOME memory, private contacts, repository data, payment authority, or live backends.

How I used Codex and GPT-5.6

Codex was my primary engineering partner: it inspected the project, implemented and debugged Swift and JavaScript, designed the MCP boundary, wrote tests, verified root-state restoration and renderer lifecycle, and prepared the public judging snapshot. GPT-5.6 powered the reasoning sessions that turned the idea into a concrete permission flow, challenged unsafe claims, and verified mission and Receipt invariants. In the recorded proof, an actual Codex CLI running GPT-5.6 calls the isolated Hubble MCP. No OpenAI API credits were required.

What I learned

An agentic social product is not credible because agents look autonomous. It becomes credible when intent, disclosure, approval, and evidence are visible. Hubble's core loop is: request, discovery, owner approval, connection, Room, Receipt.

What's next

Replace demo fixtures with consented multi-owner transport, preserve the same approval boundary, and let many AI providers enter the same world through a shared protocol.

Built With

• swiftui