Start Introduction

📖 This page mirrors the canonical Agent Karma docs. Edits should be made there, not in the wiki.

The verification gap

Why Agent Karma exists — and the one question it helps you answer.

AI writes a growing share of our code. We've gotten very good at measuring how much — lines generated, % AI, hours saved. We've gone strangely quiet about whether anyone verified it before trusting it.

That's the verification gap — and it's where "it worked in the demo" becomes a 2 a.m. incident.

Agent Karma's one question

Did you verify what the AI produced — by running tests, builds, and linters — before you trusted it?

Agent Karma is a local-first VS Code extension that reflects your validation practice back to you. It doesn't track how much AI you use; it measures whether you checked it.

It's a calm mirror, not a manager's dashboard.

Dharma → Karma → Phal

The loop is framed on an old idea: you own the action and its fruit.

• Dharma — Intent. Start with clear purpose, not blind prompts.
• Karma — Action. The AI acts; you stay in the loop.
• Phal — Outcome. Validate the result. Own what you ship.

The Bhagavad Gita put it directly (2.47): "You have the right to your action — but never to its fruit." You can run the AI. The outcome is still yours to own.

What makes it different

• Validation-first. It scores your validation discipline, objectively — not vibes, not volume.
• Radically private. 100% local. No cloud, no telemetry, no login. It never reads your source code, terminal output, or keystrokes.
• Tool-agnostic. It watches your validation actions, so it works the same whether your AI ran in Copilot, Cursor, a Claude Code terminal, or a browser tab.
• Yours. Free, open-source (Apache-2.0), and your data lives on your machine — deletable anytime.

Next: install it and read the privacy contract.