Skip to content

AI Compass / Quick reads

Whoever executes should not sign off their own work

Automation Prompting Costs

Checker and executor should be separate, even with AI agents. A small, cheap second model is an inexpensive checker, as long as it is clear what data it gets to see.

In depth: A second opinion before risky commands – yes/no guards

A Claude Code session produces many small yes/no questions: May this command run? Does the change violate a project rule? Is the task really done? AI LABS presents eight open-source tools from third parties. They hand such questions to the decision model Jev through a hook. A hook is a script that Claude Code runs automatically at fixed points. In the test, Toolgate stopped a delete command even though Claude was running without asking back, but it also blocked deletions that were intended. Abide checks changes against the CLAUDE.md. Belay sends Claude back to work if the history contains no evidence of “done”.

Tool and alternative

Jev is reachable through OpenRouter, Vercel AI Gateway or a key from TypeSafe. The video names no costs. Each call costs money at the model provider, and the amount depends on the provider. Without a model you can use fixed hook rules, such as a list of forbidden commands, or Claude Code’s built-in auto mode. Toolgate does exist: the repository RiskAverseTech/toolgate describes itself as a firewall for tool calls based on Jev, running as a Claude Code hook (checked externally on 9 October 2026). The other seven repositories have not been checked.

Sources

Advertising in the source: Video sponsored by Decodo (scraping API, 8:02-8:57), plus self-promotion for AI Labs Pro.

Want to try this with your team? That’s what the AI Lab is for. See the AI Lab