Hey Product Hunt 👋 The interesting AI agents aren't the ones that answer questions — they're the ones that do things: send money, delete data, ship code, email customers. The problem: once an agent decides to act, there's usually nothing checking whether it should. Most setups either trust the agent fully or make a human approve every step. Neither works in production. Kelhe sits in that gap. Right before an agent takes an action, it asks Kelhe — and gets back a risk score for that exact action, plus a decision: ✓ Allow — safe and routine, run it ⏸ Hold — risky or unclear, send to a human first ✕ Block — crosses a hard line, stop it Every score comes with the signals and reason behind it, so you can always answer "why did the agent do that?" It's a single HTTP request — no SDK, no framework lock-in, any language. We're early and building in the open. We'd genuinely love to hear: what are you building agents to actually do — and where are the actions that scare you? → kelhe.com · team@kelhe.com 🦊

Kelhe

Risk scoring for autonomous AI agents.

Risk scoring for autonomous AI agents.