↓Skip to main content
  1. Tools/
  2. AI Sessions/

Git AI

·3 mins·
ai-ready AI Sessions git-only Ai-Sessions Ai Attribution Git Git-Notes Blame Authorship Compliance
Language: Rust
License: MIT
Maturity: beta
Verified: 2026-03
Table of Contents

Git AI approaches the AI session recording problem from a different angle than transcript archival tools: it tracks line-level authorship – recording exactly which lines of code were written by an AI assistant versus a human developer.

Where git-memento preserves conversations and Entire.io preserves sessions, Git AI preserves attribution at the finest granularity git supports.

How It Works #

Git AI installs pre/post-edit hooks into supported AI agents. When an agent modifies a file, Git AI records a checkpoint – a small diff stored in .git/ai/ – capturing exactly what the agent changed.

On commit, these checkpoints are processed into an Authorship Log attached as a git note under refs/notes/ai. The log records, for each file in the commit, which line ranges were AI-authored and which were human-authored.

Agent Support #

Git AI supports the broadest range of AI coding agents:

  • Claude Code
  • GitHub Copilot
  • Cursor
  • Continue
  • Gemini
  • Codex
  • OpenCode
  • Droid
  • Junie
  • Rovo Dev

This breadth makes it suitable for teams where developers use different tools.

Key Features #

AI-Aware Blame #

git-ai blame path/to/file.py

A drop-in replacement for git blame that annotates each line with its authorship source (human or AI, and which AI agent). Available as CLI output and as IDE decorations in VS Code, Cursor, Windsurf, and Emacs (magit).

Attribution Survives History Rewrites #

Authorship logs stored as git notes are designed to survive rebases, squashes, cherry-picks, and amends. This is critical for teams that rebase feature branches or squash-merge pull requests.

Enterprise Dashboards #

For organizations, Git AI provides dashboards showing AI adoption metrics across repositories: percentage of AI-authored code, trends over time, and per-developer breakdowns.

How It Differs from Transcript Tools #

Git AI and transcript-based tools like git-memento or Entire.io are complementary, not competing:

Git AIgit-memento / Entire.io
What it recordsWhich lines are AI-writtenThe conversation that produced the code
Primary question answered“Who wrote this line?”“Why was this change made?”
GranularityPer-linePer-session
Use caseCompliance, blame, adoption metricsCode review, debugging, onboarding

A team could use both: Git AI for line-level attribution and compliance reporting, and git-memento or Entire.io for preserving the reasoning context.

Integration with con/serve #

For research software projects tracked with DataLad, Git AI’s authorship logs provide a provenance layer that maps naturally to research attribution concerns:

  • Which parts of an analysis pipeline were AI-generated?
  • What proportion of a software tool was human-authored?
  • Can authorship claims in publications be supported by git evidence?

The git notes storage (refs/notes/ai) integrates with DataLad the same way as git-memento’s notes – propagated via configured refspecs on push.

Agent Trace Specification #

Git AI’s approach aligns with the emerging Agent Trace specification (github.com/cursor/agent-trace), an open RFC backed by Cursor, Cognition (Devin), Cloudflare, Vercel, and Google Jules. Agent Trace defines a vendor-neutral JSON schema for AI code attribution metadata. As this spec matures, tools like Git AI may converge on a common format for interoperability.

Limitations #

  • Does not preserve conversations – authorship logs tell you what was AI-written but not why or how. Pair with a transcript tool for full context.
  • Checkpoint overhead – .git/ai/ accumulates diffs during active sessions. Long sessions with many edits may produce significant checkpoint data.
  • Enterprise features require a commercial license.

See Also #