The verification layer for
AI-assisted development.

Qmmit is developer infrastructure that automatically links AI prompts to git commits — creating verifiable proof of how software is built in the AI era.

The problem

67% of engineering job posts now mention AI tools. But there is no standard way to verify AI proficiency. Resumes claim "proficient with AI." Interviews test prompting in artificial settings. Neither reflects how a developer actually works with AI day-to-day.

What we built

A CLI that installs in 5 seconds, captures prompts from 7 AI tools via git hooks, and links each prompt to the commit it produced. No workflow changes. No plugins. No API keys.

Capture

Reads session data from Cursor, Kiro, Claude Code, Copilot, Windsurf, OpenCode, Gemini CLI.

Match

Weighted algorithm scores file overlap, timing, keywords, and code similarity. Links prompts to commits.

Verify

Every attribution is backed by a real commit SHA. Auditable. Not self-reported.

Architecture

Local-first. All data stays on your machine until you explicitly push. We never access data you have not shared.

Encrypted. AES-256 at rest. TLS 1.3 in transit. Auto-redaction catches secrets before they leave your machine.

Git-native. Standard post-commit and pre-push hooks. Works with any git workflow, any hosting platform.

Zero config. One command installs hooks, detects tools, and starts tracking. No plugins. No API keys. No proxy.

Where we're going

Qmmit captures something no other platform has: paired prompt-outcome data at scale. Every prompt linked to the code it produced. This creates a natural reinforcement learning environment.

Prompt quality scoringML models trained on prompt-outcome pairs predict effectiveness before execution.
Intelligent model routingContextual bandits learn which model performs best for each task type in your codebase.
Token efficiency optimizationReduce AI spend 30-50% while maintaining output quality.
On-chain verificationCryptographic attestations that make AI contributions independently verifiable.

Principles

Privacy is architecture, not a feature

Local-first is not marketing. It is a design constraint.

Invisible tools win

If you have to change your workflow, we failed.

Verification over claims

Anyone can say they use AI well. We provide proof.

Open standards over lock-in

Git is the interface. Any tool that saves sessions works.

Try it.

One command. Free. Takes 5 seconds.