git-ai

Executive Summary

git-ai is an open-source Git extension that records line-level provenance of AI-generated code: every AI-written line is linked to the agent, model, and prompt that produced it, with attribution stored in Git Notes and reconciled through rebase, squash, merge, and cherry-pick.^[1][2] Built in Rust by Aidan Cunniffe and Sasha Varlamov, the project launched in July 2025 and has shipped 179 releases to v1.5.6 in roughly eleven months, reaching 2,049 stars and 200 forks as of June 11, 2026.^[3][4]

The timing argument is the product. Analyses of GitClear's 211-million-line dataset estimate roughly 41% of new code is now AI-written — with no structured record of which lines, which model, or why.^[5] The April–May 2026 backlash over VS Code flipping git.addAICoAuthor on by default — Microsoft apologized and reverted after commits were silently tagged "Co-authored-by: Copilot" — proved both that developers care intensely about AI attribution and that a one-line commit trailer is too blunt an instrument for it.^[6][7] With EU AI Act Article 50 machine-readable AI-content disclosure obligations applying from August 2, 2026, git-ai is positioned as the structured, consent-by-install answer to a question regulators and platforms are now forcing.^[8] Thoughtworks placed it in the Assess ring of its Technology Radar.^[9]

---

Product Overview

git-ai installs once per machine and then manages agent hooks automatically, updating them daily — no per-repository setup.^[1] Coding agents call git-ai checkpoint as they write or modify files; each checkpoint captures a diff against the previous checkpoint, marked AI- or human-authored. At commit time the accumulated attribution is stapled to the commit as a Git Note, giving every line a durable link back to the agent, model, and prompt that generated it.^[1][9]

Day-to-day use happens through two commands: git ai blame, a drop-in replacement for git blame that shows AI attribution per line, and git ai stats, which aggregates AI-code percentages and generation-to-acceptance rates across a repo or org.^[1]

Key Capabilities

Product Surfaces

---

Technical Architecture

git-ai is a single Rust binary that extends Git rather than wrapping or replacing it. Attribution lives in Git Notes — native Git objects — so the data travels with the repository and requires no external database or service.^[1][2]

# Mac / Linux
curl -sSL https://usegitai.com/install.sh | sh

# then, in any repo with a supported agent
git ai blame <file>
git ai stats

^[2]

Key Technical Details

---

Strengths

• Right granularity for the trust debate — line-level agent/model/prompt provenance is structurally richer than the commit-trailer approach that blew up on VS Code users; attribution is opt-in at install and visible via standard Git tooling.^[1][6]
• Survives history rewrites — attribution reconciles through rebase, squash, merge, and cherry-pick, the operations that destroy naive metadata schemes.^[1]
• Broad agent coverage, zero per-repo setup — hooks for 10+ agents and editors are installed and refreshed automatically.^[1]
• Third-party validation — an Assess placement on the Thoughtworks Technology Radar, which frames it as worth exploring for "long-term accountability and maintainability in an agentic workflow."^[9]
• Standards posture — building on Git Notes and the Agent Trace spec (also implemented by Cline and OpenCode) hedges against single-vendor lock-in.^[9]
• Relentless shipping — 179 releases in roughly eleven months, with pushes through June 11, 2026.^[4][3]

---

Cautions

• The "why track at all" objection is unresolved — in the most substantive community debate on AI-code provenance (58 points, 71 comments), skeptics argued existing practice already covers intent: "We solved it already... diffs would be traceable to commits and PRs, which in turn are traceable to tickets."^[10]
• Muted launch reception — the project's own Show HN drew just 6 points and 4 comments, all friendly; community pull is not yet proven at Vouch-launch scale.^[11]
• Prompt data in repo history — storing prompts in Git Notes means session content travels with the repo; teams with sensitive prompt material need a redaction story before pushing notes to shared remotes.^[1]
• Cooperation-dependent accuracy — attribution depends on agents calling checkpoints through managed hooks; edits from unsupported tools, or Bash edits outside the repo cwd, silently fall through as human-authored.^[1]
• Measurement invites gaming — once git ai stats feeds dashboards (or mandates), AI-percentage numbers become targets; the GitClear data behind the 41% figure also shows churn and duplication rising, a reminder that volume metrics mislead.^[5]
• Hard edges — git mv, git filter-branch, and git replace are unsupported.^[1]
• 166 open issues against a two-person maintainer core as of June 2026.^[3]

---

Pricing & Licensing

Licensing model: Apache 2.0 for the core; attribution data is plain Git Notes, readable without git-ai installed.^[3][1]

Hidden costs: none at the tool level, but pushing/fetching notes refs and onboarding agents across a large org is real workflow work, and any Teams pricing is undisclosed.^[1]

---

Competitive Positioning

Direct Competitors

When to Choose git-ai Over Alternatives

• You need to know which agent, model, and prompt produced a specific line — for review, debugging, audit, or incident response — not just that "AI helped" on a commit.^[1]
• Your team runs multiple agents (Claude Code + Cursor + Copilot) and wants one attribution layer instead of per-tool trailers.^[1]
• You anticipate EU AI Act-era disclosure requirements and want machine-readable provenance accumulating now, in-repo, under an open license.^[8]
• Choose Vouch instead when the problem is unwanted contributors, not untracked code; choose Oobo when you want commit-level AI metrics for dashboards without line-level depth.

---

Ideal Customer Profile

Best fit:

• Engineering orgs with heavy multi-agent adoption that need audit-grade answers to "who wrote this line"^[1]
• Regulated or EU-exposed teams preparing for machine-readable AI-content disclosure from August 2026^[8]
• Platform teams measuring real AI contribution and acceptance rates rather than survey-based estimates^[1]

Poor fit:

• Teams whose agents or editors fall outside the supported set — unhooked edits silently degrade attribution accuracy^[1]
• Solo developers satisfied with commit messages and PR links as provenance^[10]
• Orgs that cannot allow prompt content to be stored in repository notes^[1]

---

Viability Assessment

The macro forces are unusually aligned for a one-year-old developer tool: ~41% of new code is AI-written with no structured provenance, Microsoft demonstrated both the demand for and the wrong way to do attribution, and EU disclosure rules land in August 2026.^[5][6][8] The open question is whether git-ai converts that tailwind before a platform vendor ships a good-enough native version.

---

Bottom Line

git-ai is the most technically serious answer yet to AI-code provenance: line-level, multi-agent, durable through rebase, stored in native Git objects under Apache 2.0, and validated by a Thoughtworks Radar Assess placement — all within a year of launch.^[1][9] Adoption (2,049 stars) is meaningful but not yet ecosystem-scale, community skepticism that commits-plus-tickets already suffice remains live, and accuracy depends on agents cooperating via hooks.^[3][10]

Recommended for: multi-agent engineering orgs that need auditable, line-level answers about AI authorship, and EU-exposed teams building toward August 2026 disclosure obligations.^[8]

Not recommended for: teams outside the supported agent set, orgs that can't store prompt content in repo notes, or anyone seeking a vendor-backed product with published pricing — the Teams tier is undisclosed.^[1]

Outlook: the strongest entrant in the trust-the-code slot of developer trust tooling. Its bet — that an open, Git-native provenance standard wins before platforms bundle a coarser one — looks better after the VS Code co-author debacle, but the window is regulatory: August 2026 is when structured AI attribution stops being optional for many teams.^[6][8]

---

Research by Ry Walker Research • methodology