The Built with Opus 4.7 Claude Code hackathon concluded with 500 participants across multiple countries, co-hosted with Cerebral Valley. Winners made heavy use of multi-agent orchestration, persistent memory, MCP tools, sandboxed execution, and structured prompting.
MedKit, built by a developer from Türkiye, took first place with a medical training tool where one Managed Agent plays the patient, observes the trainee's actions, and grades performance. The system prompt doubles as a citation allowlist for clinical guidelines.
Wrench Board, a French entry, placed second with a Managed Agent that pilots a boardview interface using 4-layer memory and roughly 36 tools, re-orienting across sessions by reading back its own notebook.
MaestrIA, a Chilean entry, won the Keep Thinking Prize by pairing two Managed Agents for repair diagnosis: one handles async price validation over SSE, the other writes WhatsApp outreach messages.
While one Claude instance worked on a Bun HTTP/3 server, a second Claude instance wrote a custom load testing CLI to benchmark and iterate on the server's performance. The existing tool (oha) wasn't fast enough to saturate the server, so the solution was just... more Claude.
ARIA, built by a French team, took the Best Use of Claude Managed Agents prize. It runs a single Managed Agent equipped with 17 MCP tools, using sandboxed Python to compute degradation rates.
A Danish developer's Virtual Puppet Theater won the Creative Opus 4.7 Exploration prize by running two Claude instances simultaneously: one performs the puppet character, the other procedurally generates new props from THREE.js primitives.
