AI for the parts of git you skip.

See it work

Splitting a dirty tree

$ git status
 M src/auth.py
 M src/api.py
 M tests/test_auth.py
 M README.md
?? src/rate_limit.py

$ git log --oneline
a1b2c3d feat(auth): add rate limiting
e4f5g6h test(auth): cover edge cases
i7j8k9l docs: document auth flags

Cutting a release

$ # bump every manifest by hand
$ vim package.json Cargo.toml pyproject.toml
$ # write a changelog entry
$ vim CHANGELOG.md
$ git add -A
$ git commit -m "release v0.1.0"
$ git tag v0.1.0
$ git push && git push --tags
$ # if anything failed, clean up by hand

$ imp release --minor
→ Bump: 0.0.53 → 0.1.0
→ Synced 3 manifests, wrote changelog
→ Tagged v0.1.0, pushed
# rollback on any failure is automatic

Resolving a merge conflict

<<<<<<< HEAD
rate_limit = RateLimiter(per_ip=10, window=60)
=======
rate_limit = RateLimiter(per_user=20, window=300)
>>>>>>> feat/per-user-limits

# read both branches, decide, hand-merge, repeat per file

$ imp resolve
→ src/auth.py: combine per-IP and per-user limits
  Apply? [Yes / Edit / No]
→ Done. 1 file resolved.

What it does

More examples

Commit

$ imp commit -a
→ feat: add rate limiting to API endpoints
  Use this message? [Yes / Edit / No]

Pull request

$ imp pr
→ Title: feat: rate limit auth endpoints
→ Body:  ## Summary
         Adds per-IP rate limiting to /login and /register...
  Create PR? [Yes / Edit / No]

Release

$ imp release
→ Bump: patch / minor / major
→ Squashes commits, writes CHANGELOG.md, syncs package manifests, tags v0.0.1, pushes.

Package manifests

At release time imp scans for package files (package.json, Cargo.toml, pyproject.toml, composer.json, pubspec.yaml, *.gemspec, mix.exs, build.gradle, pom.xml, Chart.yaml, manifest.json) and asks the AI which ones belong to this project. Vendored copies, fixtures, and dynamic-versioned manifests are skipped. Picked manifests are rewritten and staged into the release commit; rollback restores them on failure. Pass --no-manifests to disable.

Pre-release versions are translated to each ecosystem's native format. v0.0.52-rc.1 becomes:

imp release --rc commits the manifest bump alongside the tag so the published artifact's version matches. imp changelog --apply only labels existing commits with retroactive tags, so it leaves manifests alone.

How imp compares

aicommits and OpenCommit cover the commit-message step. imp covers the whole loop: dirty tree → clean commits → PR → release → changelog.

imp vs Claude Code

Claude Code can write commit messages too. They are different shapes for different jobs.

Use Claude Code for open-ended work where you need a thinking partner. Reach for imp for the same git ritual you do twenty times a day, where you want a verb, not a chat.

They compose: imp can use the Claude CLI as its provider, so the prompts run inside your existing Claude subscription.

Install

pip install imp-git
# or try without installing:
uvx --from imp-git imp doctor
pipx run --spec imp-git imp doctor

Configure your AI provider:

imp config
imp doctor

Works with the Claude CLI or local Ollama models. No API key required for Ollama.

Commands

Hover any token to see what it does. Type to filter.

Any AI command accepts -w / --whisper to steer the AI without overriding its rules:

imp commit -a -w "use IMP-42 as ticket"
imp review -w "focus on error handling"

Configuration

imp config writes ~/.config/imp/config.json (or $XDG_CONFIG_HOME/imp/config.json).

{
   "provider": "claude",
   "model:fast": "haiku",
   "model:smart": "sonnet"
}

Each key has a matching env var (IMP_AI_PROVIDER, IMP_AI_MODEL_FAST, IMP_AI_MODEL_SMART) that takes precedence over the file.

Requirements

• Python 3.10+
• git
• gh (optional, for fix, pr, release)
• An AI backend: the Claude CLI or a running Ollama instance