Droid ASC

Why it's ranked here

The design idea is the draw: skip the up-front indexing that most Android reverse-engineering tools insist on and answer each question from the compiled bytes. The README backs this with a Black Hat Europe Arsenal talk and its own measurement on a 352MB commercial app, a 177 millisecond class decompile in 141MB of RAM. The repository also carries a serious performance regression harness with paired, statistically gated timing comparisons in CI. Apache 2.0 and installable from PyPI. The speed figures are the author's own, measured on the author's samples.

What's good

Single-class lookups stop as soon as a match is found: parallel workers inflate code files and a shared stop signal cancels the rest. The class check is a binary search over the sorted type table, not a full parse. Output from every subcommand can go to stdout or a file, which suits scripted and agent pipelines. Testing notes describe coverage for multi-file apps, both stored and compressed entries, malformed indices and a startup budget of 100 milliseconds internal time, with CI on Python 3.11 and 3.12. The decompiled source keeps method annotations that the underlying engine would otherwise drop.

Tradeoffs

The actual Java reconstruction comes from Androguard's decompiler, which this tool imports and patches at runtime, replacing class constructors and visitor methods. That makes it sensitive to Androguard upgrades, and the declared dependency has only a lower bound. It also replaces many standard-library modules with placeholder stubs at import time to speed startup, a trick its own comments describe as fragile and partly LLM-generated. Output quality is whatever Androguard's decompiler produces, not a JADX-grade engine. There is no MCP server or library API documented in the README; the interface is the command line and a GUI.

How to use it well

Fits a reverse engineer or an automated agent that needs fast, repeated answers from large apps: list classes under a package prefix, pull one class as source, then chase string or method references across every code file. Raise the worker thread count on big multi-file apps. Keep it pinned to a known Androguard version in its own virtual environment, since the runtime patching assumes specific internals. Reach for a full decompiler when you need whole-project export, resource decoding beyond the manifest, or a polished browsing experience.

Observed

License

Apache-2.0 (LICENSE, pyproject.toml)

Language

Python, requires 3.10 or later per pyproject.toml

Install

pip install droidasc, or pip install . from source

Interfaces

CLI with getclass, listclass, getmanifest, findrefs subcommands, plus a --gui desktop mode

Core dependency

Androguard, imported and monkey-patched at runtime in droidasc/asc_core/utils/decompiler.py

Tests

Regression runner, startup and reference benchmarks, and a paired base-versus-candidate comparison described in tests/TESTING.md

Startup budget

Median 100 ms internal and 250 ms wall time for getclass --debug on Linux CI (tests/TESTING.md)

Platform notes

Fork-based worker pool on POSIX; Windows falls back to the default context (droidasc/asc_client/apk_handler.py)