TRY IT HANDS-ON functional βœ“ tested 2026-10-01
// sandboxed in macOS (host) Β· aarch64 Β·install log Β·functional log

Android Reverse Engineering Skill

by Simone Avogadro (with contributions from tajchert, philjn, txhno, muqiao215, kevinaimonster, and others) · https://github.com/SimoneAvogadro/android-reverse-engineering-skill · Apache-2.0 · vv1.5.0 (master manifests; tags run to v1.7.0) · updated 2026-09-30
Don't install your next skill blind. One email a week: what's worth installing.

The 7.9K-star Android RE skill runs its whole Phase 0 to 5 arc on a real F-Droid APK, but a vanilla macOS install breaks in three different scripts.

⚙ ⚙ ⚙ ⚙ ⚙ 3 / 5
quality 3/5
documentation 4/5
setup 3/5
value 5/5
ecosystem fit 3/5
// bottom line

The workflow design is the best in the Android RE niche: fingerprint-first triage, dual decompiler engines, and R8 Kotlin name recovery all verified on a real APK with 10,761 decompiled files and 215 recovered class names. But stock macOS breaks three of the six scripts (bash 4 syntax, a BSD awk regex, and a java-stub crash in the dependency checker), and the release line is drifted: master says 1.5.0 while tags run to v1.7.0 on a diverged branch that still ships a known install-blocking bug. Linux users get a near-4 experience today; macOS users need homebrew bash and lowered expectations.

Get the skill digest The best and the broken from 100+ hands-on tests. One email a week.
install via Claude Code marketplace
$/plugin marketplace add SimoneAvogadro/android-reverse-engineering-skill
then install the plugin
$/plugin install android-reverse-engineering@android-reverse-engineering-skill
or clone directly
$git clone --depth 1 https://github.com/SimoneAvogadro/android-reverse-engineering-skill.git

408K shallow clone

install if

  • Security researchers and authorized pentesters on Linux workstations. This is the strongest agent-native Android static-analysis workflow reviewed on GearScope, and every Linux-path probe passed on real input.
  • Developers who must reproduce a third-party Android app's HTTP API (interop work, test fixtures, migration audits). The two-tier endpoint documentation template turns a noisy 100-endpoint dump into something a human can read.
  • Analysts triaging unknown APKs at scale. Phase 0 fingerprinting in 2 seconds per app, with framework-aware redirects, saves a decompile cycle on every Flutter or RN app in the queue.

What It Does

This is a Claude Code plugin (marketplace manifest, /decompile slash command, natural-language and Chinese trigger phrases) that walks an agent through reverse engineering Android packages: fingerprint the app's framework and HTTP stack, install decompiler dependencies, decompile APK/XAPK/JAR/AAR files with jadx or Fernflower/Vineflower, analyze the manifest and package structure, recover Kotlin class names obfuscated by R8, trace call flows from UI to network, and produce a two-tier API inventory. The payload is one 313-line SKILL.md routing to 7 reference files (1,089 lines) and 11 scripts (3,031 lines: 7 bash, 4 PowerShell twins). It targets security researchers, pentesters, and developers who need to document or reproduce an Android app's HTTP APIs without source code. At review time the repo had 7,944 stars and 905 forks, Apache-2.0, 31 commits from 8 contributors, created 2026-02-02.

The Good

Phase 0 fingerprinting is the right idea and it works on real input. Instead of burning minutes decompiling a Flutter or React Native app where Java decompilation yields nothing, fingerprint.sh scans the ZIP listing and DEX strings first. On the real 12MB F-Droid client APK it printed a correct one-screen verdict in under 2 seconds: Framework: Native Android (Kotlin) with rationale kotlin_module metadata present, no Compose markers, HTTP stack: OkHttp Ktor, Serialization: kotlinx.serialization, Obfuscation: LOW. Flutter, React Native, Cordova, and Xamarin detections each carry a redirect to framework-appropriate tooling (blutter, hbctool, ILSpy), which is honest about the skill's own limits.

R8 Kotlin name recovery delivers its headline claim on real decompiled code. recover-kotlin-names.sh mines @DebugMetadata and @Metadata.d2 annotations that R8 cannot strip. On the F-Droid sources it mapped 215 classes (207 via d2, 8 via DebugMetadata) across 28 real packages, and on a planted obfuscated fixture it recovered a.b.c to com.example.app.LoginRepository while correctly leaving a metadata-free class unmapped. lookup-name.sh then resolves names in both directions and annotates grep hits with the owning class's real name: a/b/c.java:4:public final class c implements Runnable { // com.example.app.LoginRepository.

The decompile wrapper handles messy reality instead of pretending it does not exist. During our real run, jadx exited with status 3 partway through, and the script correctly reported Warning: jadx exited with status 3 after writing 10761 Java files; treating this as partial success and preserved the output. The wrapper also documents and handles XAPK split bundles, bundle-wrapper APKs with inner base.apk plus splits, config-only split skipping, and an engine-comparison mode. The dependency installer's user-local route worked exactly as documented in an isolated HOME with no sudo and no brew: jadx 1.5.6 (75MB) landed in ~/.local/share/jadx with symlinks, and install-dep.sh java on macOS without brew failed closed with manual instructions rather than pretending to succeed.

The docs are unusually honest for this niche. The README marks the PowerShell scripts as experimental community work, names each contributor with their PR number, includes a lawful-use disclaimer citing EU Directive 2009/24/EC and DMCA Β§1201(f), and its repository structure diagram matches the tree on disk file for file. The SKILL.md engine-selection table and the two-tier endpoint documentation template (flat inventory always, deep dives only for auth and payment flows) read like they were written after real engagements, not imagined ones.

The Bad

Stock macOS breaks three of the six bash scripts, and nothing in the docs warns you. macOS ships bash 3.2; fingerprint.sh (Phase 0) and find-api-calls.sh (Phase 5) need bash 4. Under /bin/bash the fingerprint script dies with line 37: ${INPUT,,}: bad substitution and, worse, exits 0, so an agent sees silent success with no output. find-api-calls.sh dies with line 91: retrofit: unbound variable (its declare -A associative array). Even under bash 5, the default scan dies on macOS at a BSD awk regex incompatibility: awk: nonterminated character class (exit 2) in the URL post-filter, so the first-party/third-party host analysis, a headline feature, never prints on macOS. decompile.sh also uses GNU find -printf in its summary, printing (none) for top-level packages on BSD find. None of bash 4, GNU coreutils, or gawk appears in the prerequisites, check-deps, or the setup guide.

check-deps.sh crashes on the macOS java stub before printing its contract. With the stock /usr/bin/java stub present (a Java binary that errors because no runtime is installed), set -euo pipefail plus java -version 2>&1 | head -1 kills the script inside the command substitution: it prints the header, then exits 1 with zero output. The SKILL.md's Phase 1 machine-readable contract (INSTALL_REQUIRED: lines that drive the auto-install loop) never materializes on a vanilla Mac. Our bash -x trace shows the script dying on the exact line that captures the stub's error message.

The release line is drifted and the latest tag ships a known install blocker. Master's manifests say 1.5.0 (bumped 2026-06-10), but the v1.x tags live on a diverged trackers-and-ads branch: the latest tag v1.7.0 was cut 2026-09-30 from a line that is 23 commits ahead of and 3 behind master. Those 3 master-only commits include two real September bugfixes (the SKILL.md name field fix for issue #26 and a portable grep fix), so v1.7.0 still ships the frontmatter without name, meaning npx skills add at the latest tag fails with missing required frontmatter field(s): name while master works. The only published GitHub release is v1.1.0 from April. Add undeclared dependencies (recover-kotlin-names.sh and lookup-name.sh require python3, which appears nowhere in prerequisites or check-deps), zero CI and zero tests across 31 commits, an error path that returns success (decompile.sh with no args prints Error: No input file specified. but exits 0), and a modest registry footprint (67 skills.sh installs for the canonical entry), and the maintenance picture lags far behind the 7.9K-star demand.

Smoke Test Results

Host-based hands-on testing on macOS 27.0 aarch64 (sbx daemon unavailable, Docker auth). Stock bash 3.2 plus homebrew bash 5.3 for the bash-4 paths, jadx 1.5.6 and Temurin JRE 17.0.20.1 installed by the skill's own installer into an isolated HOME with brew hidden from PATH, and the real 12MB F-Droid client APK as the target. Every command below ran; verdicts are the harness's own output.

Run A. Fresh sandbox, no deps preinstalled

$ git clone --depth 1 https://github.com/SimoneAvogadro/android-reverse-engineering-skill.git
βœ… 408K shallow clone
$ bash -n check-deps.sh / decompile.sh / find-api-calls.sh / fingerprint.sh / install-dep.sh / lookup-name.sh / recover-kotlin-names.sh
βœ… all 7 scripts parse
$ /bin/bash -n (stock macOS bash 3.2) all scripts
βœ… parse under 3.2 (but see runtime findings below)
$ python3 -m json.tool marketplace.json; python3 -m json.tool plugin.json
βœ… both manifests valid JSON
$ python3 (PyYAML 6.0.3) strict-parse SKILL.md frontmatter
βœ… KEYS=description,name,trigger; NAME=android-reverse-engineering
$ grep '${CLAUDE_PLUGIN_ROOT}/...' SKILL.md against plugin tree
βœ… all 17 distinct internal refs resolve on disk
$ grep every shipped reference/scripts file against SKILL.md + README
βœ… zero unmentioned files
$ diff manifest versions (marketplace.json vs plugin.json)
βœ… both agree at 1.5.0
$ curl SKILL.md at latest tag v1.7.0; grep '^name:'
❌ v1.7.0 SKILL.md is MISSING the name field: skills-CLI install at that tag fails (issue #26), fixed only on master
$ compare master manifests vs latest tag
❌ master says 1.5.0 but latest tag is v1.7.0, cut from the diverged trackers-and-ads branch (master 3 ahead / 23 behind)
$ head -3 LICENSE
βœ… Apache-2.0
$ bash fingerprint.sh --help
βœ… usage, exit 0
$ bash fingerprint.sh /nonexistent/file.apk
βœ… exit 1 with "File not found" on stderr
$ bash fingerprint.sh plain.txt (under stock bash 3.2)
❌ dies `line 37: ${INPUT,,}: bad substitution` and exits 0 (silent success)
$ bash decompile.sh (no args)
❌ prints "Error: No input file specified." + usage but exits 0
$ bash find-api-calls.sh (no args)
βœ… usage, exit 0
$ bash recover-kotlin-names.sh (no args)
βœ… usage, exit 0
$ bash lookup-name.sh (no args)
βœ… usage, exit 0
$ ls .github/workflows
❌ absent: no CI, no tests in a 31-commit, 8-contributor repo
$ grep bash-4-only constructs across scripts
❌ 2 hits: fingerprint.sh:37 ${INPUT,,} and find-api-calls.sh:91 declare -A

Pass rate: 14 of 20 checks shown. The full harness ran 28 checks and passed 22 (the seven per-script syntax checks and the manifest checks are compressed above). The failures are the release-line drift, the silent bash 3.2 breakage, an error path that returns success, and the absent test infrastructure.

Full sandbox log β†’

Run B. Sandbox with deps preinstalled

Isolated HOME, brew hidden from PATH so the installer takes its documented user-local route.

$ bash check-deps.sh (empty host)
βœ… exit 1 (required-gate fails closed)
❌ zero INSTALL_REQUIRED:* lines printed: script dies at the macOS java stub (see below)
$ bash -x check-deps.sh (trace)
❌ trace ends at `java_version_output='The operation couldn't be completed...'`, exit 1: set -e + pipefail on `java -version | head` kills Phase 1's contract on a vanilla Mac
$ bash install-dep.sh jadx
βœ… jadx 1.5.6 (75MB) to ~/.local/share/jadx + ~/.local/bin symlink, no sudo
$ ls ~/.zshrc ~/.bashrc ~/.profile after install
❌ (harness expectation) no rc write in a fresh HOME: script prints the PATH line instead; add_to_profile only appends to an existing rc file
$ bash check-deps.sh (PATH incl. ~/.local/bin, java still stub)
❌ "[OK] jadx" never reached: same java-stub crash kills the run first
$ echo $? while java missing
βœ… still exit 1, never falsely green
$ bash install-dep.sh java (macOS, no brew)
βœ… fails closed (exit 1) with manual instructions pointing at adoptium.net
$ curl Temurin 17 JRE + tar (reviewer-side)
βœ… 42M download, JAVA_HOME resolved, java 17.0.20.1 runs
$ bash check-deps.sh (full toolchain)
βœ… exit 0: Java 17 + jadx 1.5.6 OK, 4 optional tools honestly listed missing
$ jadx --version
βœ… 1.5.6

Pass rate: 8 of 11 checks shown. The full harness ran 15 checks and passed 10. The single root cause of the failures is the java-stub crash: on a stock Mac the dependency checker reports nothing installable even though its whole job is to report what to install. The install machinery itself works.

Full sandbox log β†’

Run C. Functional verification (does it do what it claims?)

Scripts run under homebrew bash 5 (the bash-3.2 story is Run A's finding). Target: the real F-Droid client APK.

$ bash fingerprint.sh F-Droid.apk
βœ… exit 0; Framework: Native Android (Kotlin); HTTP stack: OkHttp Ktor; Serialization: kotlinx.serialization; Obfuscation: LOW; recommended next step printed
$ bash fingerprint.sh garbage-bytes.apk (4KB urandom)
βœ… rejected, exit 9 (no false verdict on non-ZIP input under bash 5)
$ bash decompile.sh --no-res -o fdroid-decompiled F-Droid.apk
βœ… exit 0 in 334s; jadx status 3 handled as partial success; 10,761 .java files produced; (Top-level packages summary prints "(none)" on macOS: GNU find -printf)
$ bash find-api-calls.sh fdroid-decompiled/sources/
βœ… single-pass summary correct: OkHttp=140, Retrofit=0, HMAC/Sign=20; Retrofit/OkHttp/Ktor/Apollo/Volley sections all print
$ bash find-api-calls.sh fdroid-decompiled/sources/ >/dev/null; echo $?
❌ exit 2: BSD awk dies at `sub(/[/:]...` before First-Party/Third-Party Hosts sections ever print on macOS
$ bash find-api-calls.sh fixture/ --retrofit (planted @GET("/v1/users/profile"), @POST("v1/auth/login"))
βœ… both endpoints found as file:line:match
$ bash find-api-calls.sh fixture/ --urls
❌ dies at the same awk regex (this, not detection logic, is why planted-URL checks failed on macOS)
$ bash find-api-calls.sh fixture-clean/ (no patterns)
❌ exit 2 at the same awk step (crash, not a false positive: zero api.example.com hits printed)
$ bash recover-kotlin-names.sh fixture (@DebugMetadata c="com.example.app.LoginRepository$1")
βœ… mapping.tsv: a.b.c -> com.example.app.LoginRepository; metadata-free class correctly unmapped
$ bash lookup-name.sh mapping -o a.b.c
βœ… resolves to com.example.app.LoginRepository
$ bash lookup-name.sh mapping --grep Runnable fixture/
βœ… hit annotated: a/b/c.java:4 ... // com.example.app.LoginRepository
$ bash lookup-name.sh empty-dir/ foo
βœ… clean failure exit 1 ("no mapping.json")
$ bash recover-kotlin-names.sh fdroid-decompiled/sources/
βœ… 215 classes mapped (207 via d2, 8 via debug_meta), 28 real packages
$ bash lookup-name.sh fdroid-mapping -o <first-row>
❌ (harness assertion) first row was an identity mapping (org.acra.ErrorReporter -> itself): F-Droid is unobfuscated, so real-source rows are mostly identities; the obfuscated path was proven by the fixture above

Functional pass rate: 10 of 14 checks shown. The full functional harness ran 24 checks and passed 19.

Full functional log β†’

What the runs tell you

The workflow's claims are real: on a genuine APK the fingerprint verdict is correct, decompilation survives a partial jadx failure, Kotlin name recovery maps hundreds of classes, and the retrofit extractor finds planted endpoints with file:line precision. The failures cluster into one theme: every remaining defect is a portability or release-hygiene issue (bash 4 syntax, BSD awk, GNU find, the java-stub crash, the drifted tag line), not a flaw in the reverse-engineering methodology itself. On a Linux box with bash 4 and gawk, every functional probe here would plausibly pass.

Setup Walkthrough

  1. In Claude Code: /plugin marketplace add SimoneAvogadro/android-reverse-engineering-skill, then /plugin install android-reverse-engineering@android-reverse-engineering-skill. The plugin's slash command is /decompile path/to/app.apk.
  2. Verify dependencies with bash .../scripts/check-deps.sh. On Linux this prints INSTALL_REQUIRED: lines for anything missing; install-dep.sh then installs user-locally without sudo where possible.
  3. On macOS, before anything else: brew install bash (fingerprint.sh and find-api-calls.sh die on the stock 3.2), and expect the java-stub crash in check-deps.sh until a real JDK 17+ is on PATH. The BSD awk and GNU find gaps in find-api-calls.sh and decompile.sh remain regardless; the host-analysis sections need a Linux box or a gawk shim.
  4. Install Java 17+ and jadx (the two required deps). Vineflower and dex2jar are optional and honestly gated.
  5. Scripts also run standalone outside Claude Code; the README's manual-commands section is accurate (verified against the actual script interfaces).

Alternatives

  1. MobSF (Mobile Security Framework) - the incumbent all-in-one mobile pentest platform with static, dynamic, and API analysis behind a web UI. Far heavier to host, not agent-native, no R8 name recovery, but no bash 4 or awk portability traps.
  2. apkleaks - single-purpose CLI that runs jadx and sweeps for URLs, secrets, and endpoints. A good 80/20 substitute if all you need is endpoint extraction and you do not want a plugin.
  3. jadx CLI directly - the skill wraps jadx, fernflower, dex2jar, and grep into an agent workflow; if you already know these tools, the SKILL.md's call-flow and Kotlin-recovery references are still worth stealing.
// review provenance
reviewed by
GearScope
tested
2026-10-01 · macOS (Apple Silicon)
last verified
2026-10-01
depth
HANDS-ON
sponsorship
none, ever
// share this review
// feedback
was this review helpful?

Don't install your next skill blind.

Every week: the shortlist of skills worth installing β€” and the ones to skip β€” from 100+ hands-on tests. No spam, no affiliate links.