Naked `git-worktree` vs ait¶

ait is a Git-native version control layer for AI coding agents (Claude Code, Codex CLI, Aider, Gemini CLI, Cursor) — adding worktree isolation, attempt provenance, cross-agent memory, and reviewable promotion on top of Git. Open source (MIT), Python 3.14+, dependency-free, no SaaS, no telemetry.

If you already use git worktree add to run parallel AI agents, this page is for you. It compares the raw worktree workflow with the same workflow under ait, line for line, so you can decide whether the upgrade is worth one extra command (ait init).

Both target the same problem — kind of¶

Raw git-worktree and ait both solve isolation: keep one agent's edits out of another agent's working directory, and out of your root checkout, until you decide otherwise.

That is where the overlap ends.

git-worktree is a Git primitive. It gives you a separate directory, a separate branch, a shared object store. Nothing more. It does not know what an "AI attempt" is. It does not record which prompt produced which diff. It does not summarize what previous agents already tried. It does not have a discard verb that cleans up the working tree, the branch, the prompt log, and the captured output in one shot.

ait treats each agent run as a structured attempt: a record that links intent text, command, exit status, edited files, captured output, and the resulting commits. The worktree is one detail among many — and it is the easy detail.

For a wider walk through the design, see Why ait.

7 things you do manually with naked `git-worktree`¶

	Naked `git-worktree`	With ait
Create	`git worktree add ../repo-claude-auth -b claude/auth`	`claude ...` — wrapper provisions worktree
Name	You invent a scheme: agent + branch + date	Auto: stable attempt ID + agent + intent slug
Cleanup	`git worktree remove`, then `git branch -D`, then prune stash files	`ait attempt discard <id>` — one verb
Find a prompt	`grep` shell history, hope you saved it	`ait attempt list --intent "auth retry"`
Cross-worktree memory	None — each Claude/Codex/Aider session starts blank	`AIT_CONTEXT_FILE` summarizes prior attempts + `CLAUDE.md` / `AGENTS.md`
Promote to main	Cherry-pick or merge the branch by hand, resolve drift	`ait attempt promote <id> --to main`
Discard a bad run	Manual: remove worktree + delete branch + clear partial commits + delete stash	`ait attempt discard <id>` removes all of it

The point is not that any single column is impossible — it is that the right column collapses 7 manual steps into one verb each, and records the result for next week's you.

A concrete example¶

The raw flow for one Claude Code run, three agents in parallel:

# Three worktrees, three branches, three terminals.
git worktree add ../repo-auth   -b claude/auth
git worktree add ../repo-fixes  -b codex/fixes
git worktree add ../repo-tests  -b aider/tests

# In each terminal, remember to:
cd ../repo-auth && claude -p "Refactor the auth module"   # prompt: lost
cd ../repo-fixes && codex                                 # prompt: lost
cd ../repo-tests && aider --message "Add edge case tests" # prompt: in chat history only

# Three days later, you want to discard the auth attempt:
cd ~/code/myrepo
git worktree remove ../repo-auth
git branch -D claude/auth
# Was there a stash? An untracked file? You don't remember.

Same outcome under ait, after ait init:

# Three agents, three terminals, three attempt worktrees provisioned
# automatically. Prompts and outputs captured.
claude -p "Refactor the auth module"
codex
aider --message "Add edge case tests"

# Three days later:
ait attempt list --intent "auth"
ait attempt show 01HXY...        # full record: prompt, files, exit, commits
ait attempt discard 01HXY...     # cleans worktree, branch, metadata

The first script is something you write once, lose, and rewrite. The second script is the actual UX every day after ait init.

When you don't need ait¶

Be honest about the cases ait does not earn its slot:

One agent, one branch, ad-hoc. You spin up a single Claude Code session in a worktree once a week, throw away the branch after merge, and never go back to read what you did. Raw git-worktree is fine. The overhead of ait init is not paying for anything.
You don't run agents in parallel. If your workflow is strictly sequential — one agent at a time, in your root checkout — even raw worktrees are overkill. ait does not change that.
You never want to look up old prompts or runs. ait's biggest earned value is the queryable attempt history. If you treat AI coding as fully ephemeral and never audit it, that value evaporates.
Your repo is so small that cleanup is trivial. A 200-file repo with five contributors does not generate enough attempt churn for the bookkeeping to matter.

If three of those four describe you, stop reading and stay on raw git-worktree.

When ait pays off¶

Conversely, ait earns its keep when:

You run two or more agents in parallel, especially across Claude Code + Codex / Aider / Gemini / Cursor. The cross-agent memory layer alone repays the install.
You have stricter review discipline: you want to look at a diff with full provenance (prompt, exit status, files) before it ever touches main. ait attempt promote is the explicit verb that forces that habit.
You need audit trail for compliance, security review, or just a teammate asking "what prompt produced this diff three weeks ago?". The structured DSL beats grep over shell history every time.
You want local-only provenance — .ait/ lives next to .git/, the harness daemon is a Unix socket, no network, no telemetry, no SaaS sign-up.
You hand off between agents and don't want each new run to pay for the same investigation in tokens.

Migration: from naked worktree to ait¶

You don't need to give up the worktrees you already have. ait co-exists.

# In your existing repo (with or without active worktrees):
pipx install ait-vcs
cd your-repo
ait init
direnv allow   # only if prompted

# Verify wrappers and hooks were installed for whichever agents are on $PATH.
ait adapter list
ait doctor

# From this point, just use your agents as usual. ait detects and wraps them.
claude ...
codex ...
aider ...

Existing manual worktrees keep working — ait does not touch them. New runs through the wrapped agent CLIs become attempts. When you are comfortable, you can stop creating worktrees by hand and let ait provision them.

If anything looks wrong:

ait doctor --fix
ait repair

For the full setup walkthrough, see Getting started. For the verb cheatsheet, see the command reference.

Next step¶

If raw git-worktree is already covering your one-agent, one-branch weekly run — keep it. If you have ever lost a prompt, cleaned up five worktrees by hand, or wanted a structured "show me every Claude run that touched auth.py last month" query, run ait init once and see what changes.

→ Get started in 2 minutes

Naked git-worktree vs ait¶