Brek

AI agents now hold real credentials, and the main safeguard is a stream of permission prompts. Anthropic reports that users approve 97% of them. In one reported incident, an agent deleted a company's production database and backups in nine seconds using a token it found in an unrelated file.

Brek moves control to where people can judge it: the plan. The agent submits its plan, and AI turns it into a short contract covering what it will touch, what it won't, and what can't be undone. You approve once on your phone.

Every tool call then passes through Brek's gateway. Deterministic code, outside the model, checks each call against the contract, so the rules can't be lost when an agent's context is compacted. Routine calls run silently. A call outside the plan is held, and AI explains it on your phone. Stop revokes the agent's token at the gateway, and hold-to-run keeps irreversible steps under your thumb.

AI writes the contract and explains deviations. Code makes the final decision. In 30 self-written scripted runs, Brek held 15 of 15 deviations and let 15 of 15 routine runs through uninterrupted. In a live run, the real agent, unscripted, tried to drop a column through a plain query tool, and Brek held it. Brek covers only tools routed through it, and today it runs on a fake sandbox.

Tech stack: Groq (openai/gpt-oss-120b, openai/gpt-oss-20b), Next.js, React, TypeScript, Vitest, Upstash Redis, Vercel, Piper TTS for video narration. Code written with AI assistance from Claude.