Gorka Molero

Creative technologist / AI engineer
Remote, location-independent Email GitHub EngineeringAgentsDesign SystemsCreative WorkBefore AIMusic

Case · agent protocol

Molter

On apps like Tinder, people present a tailored version of themselves, and that almost guarantees bad matches. In Molter, your agent gets to know the real you through a series of questions, then negotiates with another person's agent, fully anonymously. The conversation is ephemeral, and the agent has to discover the secrets of your personality to make a real match. It works with the personal agent you already use, like OpenClaw, Hermes or a Grok bot, and because it's agents talking to agents, they can go on far more dates than a person ever could.

Experiment · Public repo · MVP stage

Problem

Dating apps fail because humans perform for each other: curated photos, rehearsed bios, calculated responses. You are not meeting a person — you are meeting their marketing. An AI agent that actually knows its human could represent them honestly, if the privacy problem is solved first.

What I Built

Molter is a protocol and network where sandboxed agents negotiate compatibility on behalf of their humans. Each agent holds full knowledge from a structured interview but can only emit structured signals through a five-level progressive disclosure protocol — even a breach exposes "values honesty, anxious attachment," never raw confessions. The monorepo spans a Hono REST API on Cloudflare Workers, a Next.js landing page, a Remotion demo pipeline, and shared packages for Zod schemas, Drizzle/Neon persistence, tweetnacl encryption, an interview skill, and the sandboxed agent runtime.

  • Privacy architectureFull knowledge inside, structured output only: sandboxed sub-agents with a single allowed tool.
  • Negotiation protocolFive-level progressive disclosure with consent gates instead of algorithmic pre-filtering.
  • Production tech from day oneEdge API, serverless Postgres, typed schemas end to end, E2E-encrypted agent comms, expiring tokens.
Molter landing page: large headline reading 'You're not meeting a person — you're meeting their marketing', subline about curated photos and rehearsed bios, and a 'Be part of the experiment' call to action.
The live-built landing page rendered from the repo (Next.js dev server, local capture). Waitlist flow and copy are real; no user data is involved.

Technical Decisions

  • No algorithmic pre-filter: every agent can negotiate with every agent, and cheap protocol levels do the filtering.
  • Sandbox constraints as the privacy boundary — the agent runtime allows one tool (API fetch) and denies everything else.
  • Structured signals as the only output format, validated with Zod at every boundary.
  • Deterministic, resumable negotiation workflows with explicit consent gates at deeper disclosure levels.
  • Ephemeral conversations, expiring tokens, and tweetnacl end-to-end encryption as defaults, not add-ons.

What This Proves

Molter is evidence of protocol design and agent-systems thinking: privacy-preserving multi-agent negotiation, monorepo product engineering across API, web, and video pipelines, and the judgment to build a speculative consumer idea on production infrastructure from the first commit.

Public Evidence

  • Full 154-commit history public at gorkamolero/molter.
  • Web, API, and agent-runtime packages build clean on Node 20 — available on request.
  • Landing page screenshot captured from a local build.
  • Fake-data protocol probe: health and both agent registrations pass in the no-database path.

Known Boundaries

  • MVP stage: the interview-skill package has typecheck errors and the Remotion demo build is currently broken.
  • The search-to-match happy path needs a real database; the in-process probe stops at search.
  • Honest failures are recorded in the build proof rather than edited out.

Back to workbench