An app builder that makes its agents prove their work

What it does

This open-source alpha turns conversational intent into working applications. A builder agent generates code into dedicated project repositories, while a separate read-only harness audits each completed turn against explicit, user-locked business requirements. The harness returns structured verdicts—SUCCESS, FAIL, or INCONCLUSIVE—with evidence, and re-verifies only files that changed.

The interesting bit

The builder and verifier are deliberately isolated: the harness cannot modify the repository it audits and evaluates requirements from scratch without inheriting the builder’s conversation. In a field crowded with generators, the project treats verification as a first-class subsystem rather than an afterthought.

Key highlights

• Conversational intent capture produces a product overview and explicit requirements
• Agent-generated apps live in dedicated repos with live previews and health checks
• Per-requirement SUCCESS, FAIL, or INCONCLUSIVE verdicts with evidence
• Diff-scoped re-verification runs in the background without blocking the builder
• Worker coordination includes project affinity and turn recovery
• Built on Next.js, Python, PostgreSQL, Redis, and minio; currently uses the Claude Agent SDK

Caveats

• Public alpha with several major roadmap items unchecked, including one-command self-hosting and model-provider adapters
• The harness does not yet claim “world-grounded evidence,” full production readiness, or the proposed ConvergenceBench evaluation framework
• Currently tied to the Claude Agent SDK; provider neutrality is on the roadmap but not shipped

Verdict

Worth a look if you are building or studying agentic code generation and care about verification as much as generation. Skip it if you need a polished, self-hosted product today; the project is explicit that it is early and intentionally incomplete.