Grainulator, a research sprint orchestrator for Anthropic's Claude Code, now offers an in-browser demo at grainulator.app. The tool compiles findings into adversarially challenged, confidence-graded claims with zero third-party dependencies, running entirely on Node.js built-ins.

After installing the Claude plugin marketplace via `claude plugin install grainulator`, users initiate research sprints with natural language prompts like "research how our auth system works." The system tracks findings as typed claims (factual, constraint, estimate, risk, recommendation) in claims.json, applying a 7-pass compiler to detect conflicts and biases. Confidence tiers range from ungraded "stated" claims to production-tested evidence, blocking output until conflicts resolve.

The ecosystem includes eight specialized tools: wheat for structured evidence gathering, mill for format conversion, and silo for reusable claim libraries. Enterprise deployment options support team-wide plugin management through .claude/settings.json configuration, with air-gapped deployments using pre-baked container images. All tools operate under the MIT license, eliminating supply chain risks.

grainulator.app demonstrates the concept via a PWA with mobile-first design, running WebLLM's SmolLM2-360M model locally. Features include fuzzy topic matching, progressive claim disclosure, and analytics tracking. Developers can launch autonomous research sprints via "research X using grainulator" commands, with the subagent iterating through research, challenge, and witness workflows until decision-ready confidence is achieved.

}