Agent Command Traffic Light

A deterministic switch for dangerous AI agent commands

Problem

AI agents can run terraform destroy as easily as terraform plan, and users fear one wrong call causing irreversible loss. Existing attempts at deterministic guardrails feel like hardened sandboxes, while users just want a controllable allow/block layer before production and a way to cap agent spend.

Solution

A local-first guardrail for agent commands: it parses shell and API calls an agent is about to run, allows or blocks them based on user rules (command allow list, regex, paths, spend caps), suggests alternatives when blocked, and records spend to auto-cut off over budget.

Users

Developers running local agents like Claude Code or Codex CLI; small teams adopting AI agents that need compliance controls.

MVP

CLI wrapper that intercepts agent subprocess commands; YAML allow/deny rules; spend threshold circuit breaker; local logs and alerts; no sandbox, just a pre-execution decision layer.

Revenue

Open-source core with paid team policy sync and audit logs; one-time purchase for personal use.

Why now

AI agents are moving from demos to production in 2026, making mistakes costlier, while generic sandboxes are heavy; a deterministic rule layer is a gap one developer can ship fast.