Website • Docs • Discord • Blog • YouTube • Download • Getting Started

Brokk — AI-native code platform for million-line repos

Brokk keeps LLMs on-task in large codebases with fragment-level context, code intelligence, and agentic Lutz Mode for fast, explainable edits.

Why Brokk is different

• Fragment-level context, not file blobs

  • Classes, methods, functions, summaries, exception traces, URLs, images, pasted text, issues, PRs, notes, and discard records are first-class fragments.

• No black boxes: persistent, branchable history

  • You never lose context. Every fragment and edit is persisted with undo/redo and you can branch from any past step.

• Modes for speed or control

  • EZ Mode (Lutz with auto task list execution) for hands-free flow; Advanced Mode when you want granular control.

• Agentic Lutz Mode

  • ContextAgent collects; SearchAgent expands and prunes; the Workspace becomes working memory. Keep/Forget/Note decisions are explicit and traceable.

• Built for enterprise scale

  • Dependency decompilation to source, incremental Java compiler for a tight edit-compile-test loop, BlitzForge mass refactors, and MergeAgent for conflict resolution.

Brokk Power Ranking (BPR)

The Brokk Power Ranking is an open benchmark of 93 real-world coding tasks drawn from large Java repositories.

• Public results and methodology: https://brokk.ai/power-ranking
• Measures how different models handle long, messy, real tasks with full edit-and-test loops.

In the Brokk app, a BPR meter appears above the Instructions panel (for example, "BPR: 93%") and adapts to:

• the model you selected
• the current context size for the task

This gives a quick "fitness for this task" indicator so you can choose the best cost/performance model before you start the coding loop.

How it works (20 seconds)

1. Describe your goal or question in natural language.
2. ContextAgent pulls in the minimal set of relevant fragments: classes, methods, summaries, tests, stack traces, URLs, and images.
3. SearchAgent expands and prunes the Workspace, discarding irrelevant fragments and promoting key details into Notes.
4. Lutz Mode synthesizes an outcome: an answer, a direct code edit, or a task list for complex changes that you can execute one-by-one and tweak as needed.
5. Brokk applies edits and runs builds/tests in a tight loop, feeding failures and diffs back to the model.
6. You review, commit, branch, and reuse sessions with full visibility into what the AI saw and changed.

Highlights

• Context Engineering

  • Fragment-level context: summaries, classes, methods, functions, images, markdown, URLs, stack traces, notes, and discard-context fragments.
  • URL fragments load pages as Markdown and are persisted in history with a link to the source.
  • Search Agent: symbol-, type-, and structure-aware search that understands usage patterns across the project and dependencies.
  • Exception and stack-trace fragments are enriched with metadata and linked code locations for faster root-cause navigation.
  • Structured tasks with Lutz Mode: define goals and constraints; Brokk collects, prunes, and readies the workspace with explicit Keep/Forget/Note decisions.

• Workspace & Preview

  • Workspace as working memory: Keep/Forget/Note operations with explicit, inspectable history.
  • Preview windows with search, quick edits, and promotion of selections into first-class fragments.

• Git Integration

  • Capture diffs, commits, and PRs as context; generate on-the-fly diffs for regressions.
  • First-class Git worktrees for parallel task branches and fast context switching.
  • PR review workflow: load PR diffs/comments as context, propose fixes, and push follow-up commits.
  • Automate the boring bits: optional auto-commit and auto-PR creation with change summaries; you stay in control.

• Edit Loop

  • Automatic compile/lint after code changes via an incremental Java compiler.
  • Run tests from the Workspace; failures are streamed directly back to the LLM.

• BlitzForge

  • Apply instructions across dozens or hundreds of files in parallel, with per-file context and a final holistic pass.

• MergeAgent

  • Resolve conflicts using blame-aware strategies that preserve intent from both sides with minimal manual editing.

• Issue Tracker Integration

  • Pull GitHub/Jira issues, comments, and images into structured context fragments inside Brokk.

• Dependency Intelligence

  • Decompile libraries to Java source so the model sees the exact APIs and versions used in your project.
  • Include relevant dependency types and members into the Workspace as needed so the model operates with precise API knowledge.

• Session & History

  • Sessions persist fragments, instructions, edits, and failures with undo/redo; branch from any past step to create a new session and copy workspaces as needed.
  • You can prune or compress history at any time (Keep/Forget/Note), including dropping fragments entirely, to reduce noise, improve privacy, and speed up loops.

Demos

Show quick demos:

• BlitzForge - Parallel mass refactor across a repo. Watch on YouTube: https://www.youtube.com/watch?v=i98Zzf8EN1M
• MergeAgent - Resolve conflicts with blame-aware strategies. Watch on YouTube: https://www.youtube.com/watch?v=SVmcS00urNw
• JBang - Real-world bug fix using agentic search and static analysis. Watch on YouTube: https://www.youtube.com/watch?v=t_7MqowT638
• Blender - Autonomous precision bug fix in a 21k-file C++ codebase. Watch on YouTube: https://www.youtube.com/watch?v=t2E6AOasb-0&t=4s

Who is it for?

• Enterprise staff and principal engineers working on decade-old or multi-repo services.
• Platform and developer productivity teams aiming for 5x+ effectiveness across large codebases.
• Architects and tech leads driving major modernization or migration projects.
• Power users who have hit the scaling limits of traditional AI code editors on million-line repos.

Getting Started

1. Register and get an API key

   • Sign up at https://brokk.ai and create an API key. New accounts include a $5 trial credit to try Brokk.
   • You will be prompted for this key on first launch. You can change it later in Settings.

2. Install Brokk

   • Recommended: download an installer for your OS from GitHub Releases: https://github.com/BrokkAi/brokk/releases
   • Alternative: jDeploy installers
     • macOS/Linux quick install:

       /bin/bash -c "$(curl -fsSL https://www.jdeploy.com/gh/BrokkAi/brokk/install.sh)"

     • All platforms and installers: https://www.jdeploy.com/gh/BrokkAi/brokk
     • Requires Java 21+. On Windows, use the Releases installers or the jDeploy page.

3. Launch and sign in

   • Start Brokk and paste your API key when prompted.

4. Requirements

   • For building from source: JDK 21 or newer (JetBrains Runtime recommended).
   • Supported platforms: recent macOS, Windows, and Linux distributions.

Community & Social

• Discord: https://discord.com/invite/qYQ2Zg2PX7
• YouTube: https://www.youtube.com/@Brokk_AI
• Blog: https://blog.brokk.ai
• LinkedIn: https://www.linkedin.com/company/buildwithbrokk
• X (Twitter): https://x.com/BuildWithBrokk

Contributing

For building from source, JVM heap tuning, headless executor CLI usage, contributor setup, coding standards, and the full developer workflow, see CONTRIBUTING.md.

License

Brokk is open source under the terms of the LICENSE file in this repository.