$ agitrack ┌ claude · session-1 ──────────────────────────────────┐ > add a --version flag to the cli · editing agitrack/cli.py … · added --version using importlib.metadata └──────────────────────────────────────────────────────┘ ✓ Created <aGiTrack> commit 4b2f31a. aGiTrack · Ctrl-G $ git log --oneline -2 4b2f31a <aGiTrack> add a --version flag to the cli a91c07e Initial commit $ git show 4b2f31a --stat <aGiTrack> add a --version flag to the cli agitrack/cli.py | 6 ++++++ trace + model + token cost in the body ↓
Agent code without provenance is unreviewable.
An agent edits twelve files while you get coffee; a week later nobody knows which prompt produced which change. aGiTrack makes each agent turn a commit, so review, blame, bisect, and revert just work.
<aGiTrack> add a --version flag to the cli # Interaction Trace ## User add a --version flag to the cli ## Agent Added --version using the package metadata; prints and exits. # aGiTrack Metadata commit_type: agent backend: claude model: claude-opus-4-8 session_name: session-1 context_tokens: 18204 tokens_since_last_commit_input: 411 tokens_since_last_commit_output: 2390
A live dashboard of what the agents did.
agitrack -d opens a live dashboard on localhost. Every number comes from commit metadata alone: identical on every clone, no database, no server. Explore the demo: this dashboard, on aGiTrack's own repo.
Not how much it wrote. How well you drove it.
It spots the habits that quietly cost time and tokens, correction loops, rework hotspots, ballooning context, and flags each one improving or getting worse, backed by the numbers.
It teaches you back.
Tell the learn page how much time you have. Your own agent reads your traces, spots your gaps, and writes short lessons on driving agents and on your own codebase, each with a quick check and an exercise it reviews. See real lessons the coach wrote for this repo's maintainer.
Didn't use it from day one? See the history anyway.
Built your project with an agent before aGiTrack? agitrack --backtrace reconstructs those sessions from the transcripts already on your machine, into the same dashboard.
agitrack --backtrace
agitrack --backtrace commit
Bring your own coding agent.
aGiTrack wraps the coding agent you already use. Supported today:
- ▸ Claude Code agitrack --backend claude
- ▸ OpenCode agitrack --backend opencode
Want a different backend? Request one.
Four ways to run it, your call.
TUI or background, auto or manual commits: four modes, the same per-turn trace, switch freely.
Interactive · Auto default
The agent's real TUI; every finished turn is committed from an isolated worktree. Or use with --no-worktree. Just work, get a clean traced history.
Interactive · Manual
Feels like plain git: you commit when ready, and pending turns fold in with the full AI trace.
Background · Auto
Drive the agent from any UI, desktop app, IDE, or chat window; aGiTrack tracks it headlessly and commits each turn.
Background · Manual
Headless tracking, commits on your terms. It can even start itself on git commit.
Three commands to the first traced commit.
pip install agitrack
cd your-repo && agitrack
Ctrl-G
Prompt as usual. When the turn settles, the commit is already there. Also a VS Code extension, one click, same experience.
Put your agents on the record.
Apache-2.0. Issues, ideas, and first contributions welcome.