VIBES is a flight recorder for everything AI does in your workflow. It records what your AI tools and agents did — code they wrote, decisions they made, actions they took — all stored right in git alongside your code.
VIBES tracking works in three simple steps. No cloud service, no vendor lock-in — just structured data in your repository.
Write code with Claude Code, Gemini, or Cursor — or let autonomous agents handle deployments, reviews, and operations. Nothing changes about your workflow.
Code it wrote, decisions it made, actions it took — VIBES-compatible tools automatically save audit data to a .ai-audit/ folder. No manual work required.
The audit data is committed alongside your code. It's versioned, portable, queryable, and yours to keep — forever.
Choose how much you want to track. Start simple and increase as your needs grow — each level includes everything from the level below it.
"Which AI? What action? When?"
The basics: which AI tool or agent performed the action, which model it used, and a timestamp. Think of it like a receipt. Lightweight enough to always leave on.
~200 bytes per entry — basically free
"What was it asked? What context did it have?"
Everything in Low, plus the full prompt and input context. Now you can reproduce results, audit what was asked, and trace bugs or bad decisions back to the original request.
Keeps the full prompt and inputs
"How did it reason through the decision?"
Everything in Medium, plus the AI's full chain-of-thought reasoning. For security-critical code, high-stakes agent decisions, and when you need to understand why the AI chose a specific path.
Full reasoning trace included
VIBES isn't just about record-keeping — it's about being prepared for real scenarios your team will face, whether AI is writing code or making operational decisions.
A vulnerability is discovered in a specific AI model. With VIBES, you can find every line it wrote and every decision it made across your entire system in seconds — instead of guessing what's affected.
An autonomous agent approves a risky deployment or makes a poor scaling choice. VIBES lets you trace back exactly what it was asked, what context it had, and why it chose that path — so you can fix the root cause, not just the symptoms.
Your organization needs to know what's AI-driven — code and operations. VIBES answers that instantly, for both code generation and agent decisions. No manual annotation, no guesswork, no auditor headaches.
Team leads and architects want to understand AI's role in your workflow. VIBES makes AI involvement visible and measurable, so you can make informed decisions about where to apply human review.
Here's what VIBES looks like in practice — for both AI-assisted coding and autonomous agent operations.
"Write a function that validates email addresses and returns a boolean"
It writes validateEmail() in your utils/validation.js file. So far, this is your normal workflow — nothing different.
A new entry appears in .ai-audit/ noting: Claude Code, claude-sonnet-4-6, the prompt you gave, and the files it touched. All automatic.
A security advisory mentions a model vulnerability. You run vibecheck query --model claude-sonnet-4-6 and instantly find every file that model touched. Remediation starts in minutes, not days.
Your CI/CD pipeline includes an agent that reviews PRs for security issues and auto-approves if it looks clean.
The agent's action is logged: which PR, what it analyzed, its assessment, and the approval decision. The full reasoning chain is captured at the High assurance level.
Weeks later, a security issue is found in that PR. You trace back to the agent's VIBES record and see exactly what it considered, what it missed, and why it approved — so you can improve the agent's instructions.
Ready to know what AI wrote in your codebase? Here's how to get started.
Start capturing what AI does — every line of code and every decision. Add VIBES tracking to your tool in minutes.
Install for your tool →Ask your AI tool provider if they record what the AI actually did. Transparent tracking should be the default, not an add-on.
How to ask →All three assurance levels — Low, Medium, and High — captured automatically. Full VIBES tracking out of the box.
runmaestro.ai →