Brokk.ai Open Source AI IDE for large codebases

Brokk

Brokk: Under the Hood

1. Introduction

• Brokk is an open source IDE designed for supervising AI coders.
• Focuses on context management for long-form coding in English, not just tab-completion.
• GitHub repository.

---

2. Quick Context: JLama, MiniLM-L6-v2, and Gemini 2.0 Flash Lite

• Quick Context Suggestions
  • Shown as blue suggestions while typing/dictating instructions.
  • Can be added to Workspace via right-click.
• Latency Minimization
  • Uses Gemini 2.0 Flash Lite for speed.
  • GPT 4.1 nano tested but found slower.
  • 400ms debounce is standard but not ideal for programmers’ workflow.
  • High frequency of calls led to auto-blacklisting by Gemini.
• JLama Integration
  • Java-based inference engine (like llama.cpp).
  • Uses MiniLM-L6-v2 for semantic embeddings (chosen for speed and size).
  • JLama checks if new instructions are semantically distinct before making LLM requests.

---

3. Deep Scan and Agentic Search: Brute Force and Tool Calls

• Deep Scan
  • Suggests additional files needed for a task using brute force (entire summarized project).
  • Uses a smarter, slower Edit model LLM; run only on user request.
  • Provides more accurate recommendations than Quick Context.
  • Recommends whether to edit or summarize files.
• Quick Context vs Deep Scan
  • Both are single-turn inference (one-shot recommendations).
• Agentic Search
  • Multi-turn process where LLM uses tools to explore codebase.
  • Useful when single-turn methods are insufficient (e.g., large projects, missing details).
  • Slower but yields high-quality results.

---

4. Code Intelligence: Joern and Tree-sitter

• Joern as Code Intelligence Engine
  • Chosen after evaluating alternatives (CodeQL, SciTools, SonarSource, SCIP, Semgrep, Tree Sitter, LSP).
  • Criteria: OSS, no special build integration, type inference, speed on large codebases.
  • Downsides: JVM-based (Scala), less uniform API across languages.
• Tree-sitter Integration
  • Added in Brokk 0.9 for partial support of non-Java languages (Python, JavaScript, C#).
  • Used mainly for summarization to reduce token usage (~10x reduction).
  • Utilizes tree-sitter-ng Java wrapper.

---

5. Conclusion / Wrapping Up

• Brokk empowers users to supervise AI coding while handling editing/syntax details.
• JLama ensures fast suggestions; Joern provides codebase insight; open-source nature allows extensibility.
• Designed for large enterprise codebases, not just demos or small projects.
• Try Brokk.

---

6. Footer / Additional Links

• Product: Home, Pricing
• Resources: Blog
• Legal: Privacy Policy, Terms of Service

---