Hey Product Hunt 👋 I built Marginal around a question I kept coming back to: When should software actually be trusted to stop an AI coding agent? Agents sometimes repeat actions without making progress. The obvious solution is to add guardrails and block repetition, but that creates another problem: repetition can be intentional, useful, or necessary. So Marginal takes a different approach. It starts in Shadow Mode and has no enforcement authority. It observes actions, outcomes, workspace state and evidence. Only after collecting and verifying enough local evidence can it earn narrowly scoped authority to intervene. Some of the ideas behind it: → Shadow Mode before enforcement → Earned Enforcement instead of static blocking → Hash-chained Decision Ledger → Authority can be automatically removed when evidence degrades → Local-first and provider-neutral → Zero mandatory runtime dependencies Current integrations include Codex, Claude Code, OpenCode and PrivacyCode. Codex currently supports limited native tool enforcement; the others are observation-only. I'm also deliberately trying to avoid inflated AI benchmark claims. In one exploratory paired Codex experiment we observed 24.93% fewer effective tokens, but neither lane solved the tasks and Marginal issued zero denials — so we explicitly do not claim Marginal caused that reduction. It's early, open source, and I'm looking for people interested in coding agents, evaluation, observability and AI governance. I'd particularly love feedback on one question: What evidence would you require before trusting a system to block an AI agent action?

Marginal

Governance that AI coding agents have to earn

Governance that AI coding agents have to earn