Agent Karma is an open-source, local-first VS Code extension that helps you close your verification gap — turning "did I actually check what the AI wrote before I trusted it?" into a visible, objective habit. It works with Claude Code, GitHub Copilot, Cursor, ChatGPT, and others — including browser/copy-paste workflows.
It is not another AI coding assistant. It is not an enterprise analytics dashboard. It is not a surveillance tool.
New to the idea? Start in the Agent Karma web app: walk through the concept, then practice in the Validation Dojo — real scenarios scored by the same engine the extension uses. It's an installable PWA (add to your home screen, works offline). Build the verification habit in the app, then bring it into VS Code.
📱 Scan to learn & experiment on your phone — or visit ai-first-community.github.io/Agent-Karma/learn
The problem with AI code isn't generating it — it's trusting it. The industry now has a name and a number for this:
96% of developers don't fully trust AI-generated code — yet only 48% always verify it before committing. (Sonar, State of Code 2026)
AWS's Werner Vogels calls the growing pile of unchecked AI code "verification debt." As agents write more of our code (already ~42% of committed code, heading toward 65%), the developer becomes the last line of trust — and most of us aren't reliably holding that line.
The industry's answer is to automate the review (Anthropic Code Review, CodeRabbit, Sonar…) — machines checking machine code. Agent Karma's answer is the opposite: keep the human in the loop by making verification a visible habit, without automating your judgment away — and without surveillance, gamification, or your code ever leaving your machine.
Every other tool measures AI usage for managers (acceptance rates, lines, tokens). Agent Karma measures one thing, for you: did you validate the AI's work? It is the mirror, not the dashboard.
Agent Karma is built on a simple idea: every action has a consequence.
| Concept | Meaning | In Agent Karma |
|---|---|---|
| Dharma | Intent, purpose, direction | What you asked the AI to do |
| Karma | Action | The changes you and the AI actually made |
| Phal | Outcome, fruit, consequence | Whether the result is validated and ready |
The thesis in one line: unvalidated Karma bears uncertain Phal. An action you never verified produces a fruit you can't trust. Agent Karma helps you notice the difference.
For each AI-assisted coding session, Agent Karma produces:
- 🪔 Dharma Card — your intent, prompt clarity, expected validation, and risk level
- 🔗 Karma Trace — a chronological, privacy-safe timeline of what happened (files saved, validation commands run, git diff summary)
- 🍃 Phal Card — the outcome: files changed, tests/build/lint detected, and whether it's ready for commit or review
- ⚖️ Karma Score — an objective, transparent score built only from the validation actions you actually took (tests/build/lint run, test coverage, change measured) — never a vague self-rating. Every point is explained.
- 🩺 Validation Context Health — "can you even validate?" — a config-only scan of your workspace for the means to verify AI output (test/build/lint/type check, a pre-commit net, CI, and whether your CLAUDE.md/AGENTS.md asks the AI to validate); it names your biggest gap and offers a one-click fix
- 🛡️ Pre-commit nudge (opt-in) — a local git hook that reminds you to validate AI-assisted changes before you commit them
- 📊 Insight dashboard — a calm, theme-adaptive view: a 🛞 karmic reflection, validation-consistency strip, Karma & validation trend lines, a task×check heatmap, risk×validation alignment, habit trends, a high-risk watchlist, and "what your Karma is made of"
- 🧾 Local AI usage (opt-in, Claude Code) — reads Claude Code's local session logs (no network, no API key, metadata only) to show what your AI work cost — tokens, turns, plus wastage (tokens spent on unvalidated work)
- 💬
@agentkarmachat participant —/verify(logs a validation — covers browser & copy-paste AI) and/summary - 🏅 Shareable Karma Card — a personalised, self-explanatory certificate of your validation practice; export as SVG or print to PDF, generated entirely locally
- 📤 Export — your session as JSON or Markdown
- 🗑️ Delete everything / Reset history — wipe all local data, or just clear your Karma history while keeping settings
Everything is stored as plain JSON on your machine.
Agent Karma is the only tool that combines all of these:
- ✅ Validation-first — it measures whether you verified the AI's output (the actions you took: tests/build/lint/coverage), not how much code it generated. No competitor centers this.
- ✅ Radically private — no cloud, no login, no telemetry, no source upload, no terminal-output capture, no keystroke/scroll surveillance. This is the part incumbents structurally won't copy.
- ✅ Coaching, not judgment — objective, self-comparative, transparent, encouraging. No leaderboards.
- ✅ Developer-owned — readable local JSON, full export, one-click delete.
And it's tool-agnostic by nature: it works the same whether your AI ran in Copilot, Cursor, a Claude Code terminal, or a browser ChatGPT tab — because it watches your validation actions, which don't care where the code came from.
See docs/differentiation.md for the full comparison against GitHub Copilot Metrics, Microsoft's AI-Engineering-Coach, CodePause, Git AI, WakaTime, and more.
Local-first. No source code captured. No terminal output captured.
No cloud upload. No telemetry. No login. No surveillance.
What it does record is metadata only: file names (including edits made by AI agents or the CLI, gated by a setting), the validation command types you ran, git diff counts, and short commit SHAs — never file contents or terminal output. Your typed intent text is recorded locally by default so your cards read back meaningfully; turn it off any time with agentKarma.capturePromptText (either way it never leaves your machine).
Read the full contract in PRIVACY.md.
🚧 Pre-1.0, feature-complete and stable. The full experience is built and tested — sessions, objective Karma Score, the insight dashboard, the opt-in pre-commit nudge, validation context health, the chat participant, opt-in local AI usage, and the shareable Karma Card. Every release ships with unit + integration tests, a zero-runtime-dependency build, and a CI-enforced no-network guard. Polishing toward a 1.0 Marketplace listing. See CHANGELOG.md and docs/roadmap.md.
| Doc | What's in it |
|---|---|
docs/vision.md |
The broad vision + full Question Map (every question, by horizon and rigor) |
docs/product-strategy.md |
Positioning, philosophy, locked decisions, non-goals |
docs/differentiation.md |
USPs and full competitive comparison |
docs/competitive-coverage.md |
Per-feature verdict on every competitor capability (Adopt / Adapt / Reject) |
docs/specification.md |
Functional spec — sessions, cards, commands, capture, testing |
docs/architecture.md |
Architecture, folder structure, data model |
docs/scoring-model.md |
Karma Score, prompt hygiene hint, Dharma/Phal generation |
docs/roadmap.md |
Phase-wise release plan with acceptance criteria |
docs/implementation-plan.md |
Task-by-task build sequence (foundation-first) |
CONTRIBUTING.md |
Build rules & phase-wise protocol for contributors |
Apache-2.0 — see LICENSE. A purely individual, community contribution. Not affiliated with any employer or vendor.
Third-party components (including the Manrope font under the SIL Open Font License) are credited in THIRD-PARTY-NOTICES.md.
Disclaimer. Agent Karma is provided "AS IS", without warranty of any kind, express or implied. To the fullest extent permitted by law, the maintainer is not liable for any damages arising from its use (see Apache-2.0, Sections 7–8). It is a habit and awareness aid — not a guarantee of code correctness, security, or production-readiness. You remain responsible for validating and shipping your code.
Use any AI coding tool. Agent Karma helps you use it better.