Built for the AI era

AI agents work fast.
Your Git history shouldn't suffer for it.

Cursor, Claude Code, and Copilot generate dozens of changes across your codebase in minutes. Prume splits them into clean, focused commits — automatically.

Download Prume Pro
pip install prumeFree forever. MIT.

Has this ever happened to you?

~/your-project

You fire up Cursor. You ask it to:

  • - Add a new /users API endpoint
  • - Fix that CSS bug on the settings page
  • - Update the README
  • - Refactor the auth middleware

20 minutes later, the agent is done. 47 files changed.

$ git status

Changes not staged for commit:

  modified: src/api/users.py

  modified: src/api/auth/middleware.py

  modified: src/components/Settings.css

  ... and 44 more files

Now what? One giant commit? 45 minutes of git add -p?

There's a better way.

Your AI agent writes the code.
Prume organizes the commits.

Works with every AI coding tool — Cursor, Claude Code, Windsurf, Copilot Workspace, or any other agent.

Messy Changes

47 files changed

Analyze

14 change units

Classify

4 groups

Clean Commits

4 focused PRs

Analyze

01

Prume scans your working directory and extracts every individual change unit (hunk) from every file.

Classify

02

Multi-layer rules — file paths, keywords, regex, then optional AI — sort changes into groups by concern. Most changes classified instantly.

Commit

03

One command creates clean, focused commits for each group. Jira issues update automatically. Stacked PRs created.

Predictable by Default. AI When You Want It.

Prume's classification engine is deterministic rules first. AI handles only the edge cases.

47 change units enter
Layer 1: File Paths60%

src/auth/** → auth-refactor, *.css → styling, tests/** → tests

deterministic60% total
Layer 2: Content Matchers20%

Diff contains TODO/FIXME → cleanup, adds import from newlib → migration

deterministic80% total
Layer 3: Keyword & Regex10%

Custom rules you define once, applied every time

deterministic90% total
Layer 4: AI Classification10%

For ambiguous changes the rules can’t catch

optional · Pro100% total
All changes classified

The majority of changes are classified by layers 1–3 — fast, predictable, no API calls, no surprises. AI handles the edge cases.

“Can't the AI just organize my commits?”

Fair question. We tried every approach. Here's what we found.

ChallengeAI Agent AlonePrume Pro
Context windowDrops context at 50+ files, hallucinating groupingsWorks at file-system level — no context window limit
Speed30-60+ seconds per large changesetRule engine classifies in milliseconds
ReliabilityNon-deterministic — different groupings each timeSame rules → same groups, every time
Silent errorsPlausible-looking mistakes you can’t catchWeb UI shows every hunk, every group, every diff
Cost$2-4 per session for large diffsRule-based classification is free; AI is targeted
LearningStarts from zero every sessionRules persist in .prume/config.yaml, improving over time

PRs that actually get merged

The #1 reason PRs get sent back: too many unrelated changes in one PR. Prume fixes that.

Before Prume

PR #142

“Add user endpoint”

47 files changed

1,200 lines added

340 lines removed

Reviewer: “Please split this up”

After Prume

PR #142Add /users endpoint
Approved

8 files · 120 lines

PR #143Refactor auth module
Approved

12 files · 200 lines

PR #144Fix settings page CSS
Approved

3 files · 25 lines

PR #145Update README
Approved

2 files · 15 lines

Built for massive changesets

Prume was forged during the Apache Unomi 3 rewrite — a multi-tenancy overhaul spanning hundreds of files across dozens of concerns. Traditional approaches and AI-only solutions both failed at this scale. That's why Prume's hybrid rules-first architecture exists.

Read the full story

See it in action

From messy working directory to clean, focused commits — in seconds.

~/acme-webapp

$ prume analyze

  Scanned 47 files, extracted 14 change units

$ prume classify

  12 change units classified by rules (layers 1–3)

  2 change units classified by AI (layer 4)

  → 4 groups created

$ prume status

Add user management API (ACME-247)

5 files · 5 hunks — pending

Fix settings page layout (ACME-312)

2 files · 2 hunks — staged

Update documentation

3 files · 3 hunks — committed

Refactor auth middleware (ACME-290)

2 files · 2 hunks — pending

$ prume commit --all

  ✓ Created 4 focused commits

  ✓ Updated 3 Jira issues

Prume Pro

Prefer a visual interface? Prume Pro adds a Web UI for drag-and-drop grouping.

Prume Pro
Prume Pro dashboard showing project overview with groups and change units
Dashboard — project overview at a glance
Prume Pro
Prume Pro groups view showing auto-classified change groups
Groups — changes classified by concern

Frequently asked questions

Everything you need to know about Prume.

What is Prume?+

Prume is a developer tool that organizes messy Git changes into clean, single-concern commits. It uses a 4-layer classification pipeline with predictable, deterministic rules and optional AI to split tangled working directories into focused, atomic commits.

Does Prume send my code to external servers?+

No. The free CLI runs entirely locally. Prume Pro's AI classification sends only diff content to the configured LLM provider, never full source files. Your source code never leaves your machine for any other purpose.

How is Prume different from interactive rebase?+

Interactive rebase (git rebase -i) requires you to manually split commits after they're created. Prume works before the commit — it classifies individual hunks into groups and creates clean commits from the start, with no history rewriting needed.

What languages and frameworks does Prume support?+

Prume is language-agnostic. It works at the Git diff level, so it supports any programming language, framework, or file format that Git can track.

Can I use Prume without the AI features?+

Yes. The free CLI uses only deterministic rules — file path matching and manual grouping. AI classification is a Pro feature and completely optional. Most changes are classified by rules alone.

How much does Prume cost?+

The Prume CLI is free forever and MIT-licensed. Prume Pro Individual costs $8/month or $79/year. Team plans start at $7/seat/month billed annually. Enterprise pricing is custom.

Does Prume work with my existing Git workflow?+

Yes. Prume operates on your working directory and staged changes and creates standard Git commits. It doesn't modify your branching strategy, CI/CD, or remote repositories.

How do stacked PRs work in Prume?+

Prume Pro can export commit groups as stacked pull requests. Each group becomes a separate PR with proper dependency ordering. Prume integrates with Graphite and GitHub's native stacked PR support.

Ready to bring order to the chaos?

Try the free CLI or get the full Prume Pro experience with a 14-day free trial.

pip install prume

Free forever. MIT licensed.

Download Prume Pro

14-day free trial. No signup required.