ContextCraft

ContextCraft is an open-source Python tool, published under the MIT License, that compiles a Git repository into a Markdown, Claude XML, or JSON context artifact sized to a chosen token budget such as 4k, 16k, 32k, or 128k. It is available as a command-line tool, a Python SDK, and a local web playground. The latest release listed on GitHub is v0.3.1.

What It Is

ContextCraft is a codebase compressor for LLM prompts. Rather than concatenating every file verbatim, it parses source with Tree-sitter, ranks files by architectural importance, and decides how much detail each file gets so the output fits a target token budget. The README positions it against repository packers such as repomix, gitingest, and code2prompt.

How the Compression Pipeline Works

The README describes a five-stage pipeline:

• A repository scanner honors hierarchical .gitignore and .contextignore files, filters binary and oversized files, and a sanitizer scrubs API keys, PEM blocks, and high-entropy secrets before parsing.
• A registry of 35 languages and formats produces ASTs with symbols, docstrings, and a lexical scope graph, and unused imports are trimmed in signature mode.
• A dependency-graph scorer combines Personalized PageRank with BM25F, biased by an optional focus query, to rank files.
• Each file is rendered at one of four fidelity tiers: FULL, SKEL (signatures, types, and docstrings), SIGS (public declarations), or TREE (a directory-tree entry).
• A multiple-choice knapsack optimizer allocates tiers by marginal utility density, which keeps very large files from consuming the full-detail budget.

Usage Paths

The CLI runs via uvx or after pip install contextcraft-ai, with options for budget, focus query, output format, clipboard copy, tokenizer encoding, and include/exclude patterns. The Python SDK can analyze a repository once and then render several budgets or formats from the same snapshot without reparsing. A FastAPI-based web playground lets users tune budgets, inspect PageRank scores, and compress public GitHub repositories by URL.

Benchmark Claims

The README reports a reproducible benchmark on its own repository (258 files, 119,813 raw tokens), with an 8k budget emitting 7,991 tokens at a stated 14.99x compression ratio. The README also recommends pairing the output with local models for offline workflows.