Harden – 5 AI models audit your code, then debate each other's findings

I built harden because I kept copy-pasting code between ChatGPT, Claude, and Gemini trying to cross-check their reviews. Each one found things the others missed, but synthesizing their outputs manually was painful.

harden runs 5 frontier models (Claude, GPT-4o, Gemini, Mistral, DeepSeek) in parallel on the same input. They analyze independently, then cross-examine each other's findings. A coordinator synthesizes the debate into consensus findings and produces a fixed version.

The key insight: no single model finds more than ~72% of issues. The union of all five hits ~94%. After cross-examination (where models must defend findings against skeptical peers), accuracy rises to ~97% and false positives drop ~60%.

How it works: - Round 1: All 5 models audit independently (no groupthink) - Debate: Each model reviews others' findings, provides evidence for/against - Consolidation: Only findings that survive cross-examination make the report - Fix: Coordinator produces a revised version addressing consensus issues - Round 2+: Same pipeline runs on the fixed version, catching fix-introduced bugs

Started with smart contract audits but it generalizes — legal docs, resumes, fact-checking, financial analysis all benefit from multi-model consensus.

Free tier available.

Built with React, Node, SSE streaming for real-time progress. The debate transcripts are the most interesting part — watching GPT-4o argue with Claude about whether a reentrancy vector is exploitable is genuinely useful.

Blog with more details on the multi-model approach: https://harden.center/blog