jungle on2.9Kpeak working-set tokenson the rich repo, vs 176K for the plain agent — flat ~1–3K at any repo size or session length
9.5×on single taskstoken reduction on batching-friendly suites
0hallucinated editsacross all benchmark runs — the edit gate only accepts edits to text the model was shown
90+languages indexedtree-sitter harness; 25K-file repos index in minutes, crash-resumable
How it works
Turn it on. Keep coding.
- IndexTree-sitter outlines and local embeddings over 90+ languages, per-file in mixed-language repos. Model-free, crash-resumable, and fast enough for 25K-file repositories. No repository content leaves the box.
- CompileEach turn gets a freshly assembled working set: outlines, expanded symbols, pinned edit targets. Context stays around 1–3K tokens however long the session runs.
- GateThe edit gate is structurally identical to the renderer, so it accepts an edit only when the model has actually seen that text. That construction is why benchmark runs record zero hallucinated edits.
Install
One install, one command.
The driver is auto-detected: a working claude binary, then ANTHROPIC_API_KEY, then a local Ollama server. Override with --driver. Python 3.11 or newer.
A Python wheel — pip install the downloaded file.
$
pip install jungle_brain-0.1.0-py3-none-any.whl$
cd your/repo$
jungleResearch
The paper behind Jungle.
Research & architecture strategy report
Jungle Memory Hierarchy
An overview of the key mechanisms
An exploration into how economic thought and models can be technically applied to make emerging technologies more accessible.