English | 简体中文

ddc decompiles Android DEX bytecode back into readable Java — at real-world app scale, queryable like a database, and javac-verified: every one of the 909,689 files it emits across seven real-world APKs parses cleanly.

Highlights

• Fast — a 226 MB / 20-dex APK (weibo, 98k classes) fully decompiles in ~21s at high quality (~11s with --fast), a 62 MB app (Telegram) in ~6.9s (~4.1s with --fast); pathological classes run on deadline-bounded monitored threads instead of hanging the run.
• Compiles — the full output of all seven benchmark APKs (reqable, Telegram, WhatsApp, weibo, weixin, qq, lark — 0.91M files) passes javac with zero syntax errors; every class of a case-variant name pair (X/Cua vs X/cua) is preserved as its own file instead of the last one silently overwriting.
• Progressive decompilation — 15 query subcommands (strings, cross-references, hierarchies, manifest, resources, per-method decompiles) answer in milliseconds: query metadata first, decompile on demand, skip the full run entirely.
• DEX 035–041, complete — multi-dex APKs, XAPK/APKS/APKM containers, invoke-custom; lambdas and string concats fold back into real Java ((a) -> …, Cls::name, a + b).
• Readable output — jadx-style local names (str, getName() → name, Kotlin Intrinsics parameter strings) instead of v12; IntDef literals render as their constants out of the box (setVisibility(8) → View.GONE; a built-in android-37 domain table, --symbols overrides); Kotlin null-check noise is elided and synthetic access$NNN bridges inline at their call sites.
• Reproducible output — no timestamps; two runs diff cleanly.
• Human-shaped output, by default (cure): every file goes through a semantic-preserving source-level pass that inlines register-copy chains, folds if/else assignments into ternaries, recovers for-each, drops dead stores and compiler artifacts — −15% output lines on real corpora with identical javac-gate results. The pass is exact, so it costs time: weibo (110k files) takes 21 s optimized vs 10 s with --fast (skip the simplification, keep formatting).
• HTTP REST + MCP, built in (xyz-rust): ddc serve exposes every query as a REST route (GET with query params or POST with a JSON body) plus GET /openapi.json and a streamable MCP endpoint at /mcp — point an MCP client at http://addr/mcp; ddc http is REST-only (no /mcp); ddc mcp stdio|http runs the same commands as MCP tools with input and output schemas. Query commands take --format auto|text|table|markdown|md|json|jsonl (auto: table on a terminal, the stable text form when piped).
• Bilingual CLI — messages localize automatically (DDC_LANG=zh|en forces; English fallback).
• Built on jdc-core, the machine-neutral core shared with jcdc.

Install

macOS (Homebrew):

brew install ejfkdev/tap/ddc

Windows (Scoop) — add the bucket once, then install by name (scoop does not resolve a bare user/repo/app form — issue #3):

scoop bucket add ejfkdev https://github.com/ejfkdev/scoop-bucket
scoop install ddc

# or install straight from the manifest URL, no bucket needed:
scoop install https://raw.githubusercontent.com/ejfkdev/scoop-bucket/main/bucket/ddc.json

cargo-binstall (any platform — fetches the prebuilt release binary instead of compiling):

cargo binstall ddc-cli

Binaries: every release ships raw binaries (no archives) for linux-amd64, linux-arm64, windows-amd64, windows-arm64, macos-amd64 and macos-arm64 — non-macOS builds UPX-compressed.

From source (Rust stable):

git clone https://github.com/ejfkdev/ddc && cd ddc
cargo build --release

Usage

ddc app.apk                          # full decompile → app-out/
ddc app.apk -c com.example.Foo       # one class to stdout
ddc app.apk -o - | less              # everything to stdout

ddc info app.apk                      # app context (label via arsc, package,
                                      # version, md5) + per-dex counts
ddc listclasses app.apk              # package tree with class counts
                                      # (--full: raw names)
ddc mainactivity app.apk             # entry point: package + launcher
ddc findrefs app.apk string token    # every const-string "token" site
ddc getmethod app.apk Foo.toString   # one method, all overloads
ddc pkg app.apk --app -o own/        # the app's own code only

Full options and examples: ddc --help (subcommand menu grouped by workflow). The complete reference — every option's semantics, every subcommand's output format and behavior — lives in docs/cli.md (中文).

Performance

Seven real-world APKs, release build, measured after a cooldown (Apple Silicon, 6P+12E). The 39-APK validation corpus (4.08M decompiled files, per-APK wall time / peak RSS / javac parse gate) is in docs/validation.md (中文).

Full decompile — 7 real-world APKs

APK	Size	Full decompile	--fast	Peak RSS (full)
reqable	34 MB	0.83s	0.50s	280 MB
Telegram	62 MB	6.91s	4.08s	924 MB
WhatsApp	139 MB	24.98s	11.41s	1067 MB
weibo	226 MB	21.03s	10.87s	1380 MB
weixin	268 MB	40.21s	25.29s	1497 MB
lark	398 MB	26.29s	14.61s	2406 MB
qq	374 MB	57.00s	31.96s	2511 MB

Recorded 2026-10-22 with v0.1.33: one run per cell after a short settle; the previous table was recorded on an earlier (pre-cure) version; full and --fast are now separate columns. Peak RSS is /usr/bin/time -l.

Query subcommands on the same APKs (cells: time / peak RSS; strings -f <package> --with-locations, findrefs on the package string and method onCreate, hierarchy/disasm/getclass on each app's launcher class):

Query subcommands — 13 commands × 7 APKs

Command	reqable	Telegram	WhatsApp	weibo	weixin	lark	qq
info	0.24s / 55MB	0.41s / 117MB	0.85s / 316MB	1.32s / 569MB	1.61s / 602MB	2.46s / 969MB	2.30s / 1145MB
listclasses	0.02s / 25MB	0.03s / 89MB	0.04s / 96MB	0.14s / 337MB	0.16s / 371MB	0.18s / 251MB	0.24s / 311MB
manifest	0.00s / 12MB	0.00s / 19MB	0.00s / 26MB	0.01s / 48MB	0.01s / 37MB	0.01s / 34MB	0.01s / 54MB
mainactivity	0.02s / 20MB	0.06s / 72MB	0.08s / 136MB	0.17s / 199MB	0.18s / 199MB	0.22s / 226MB	0.31s / 300MB
res	0.00s / 12MB	0.00s / 20MB	0.00s / 27MB	0.02s / 56MB	0.01s / 37MB	0.01s / 36MB	0.01s / 57MB
largest	0.02s / 28MB	0.05s / 71MB	0.13s / 195MB	0.32s / 363MB	0.30s / 346MB	0.46s / 519MB	0.66s / 631MB
strings	0.02s / 20MB	0.04s / 55MB	0.12s / 117MB	0.23s / 177MB	0.28s / 198MB	0.18s / 192MB	0.49s / 295MB
findrefs-string	0.02s / 31MB	0.03s / 74MB	0.04s / 203MB	0.06s / 397MB	0.08s / 425MB	0.04s / 322MB	0.10s / 849MB
findrefs-method	0.02s / 31MB	0.05s / 81MB	0.04s / 212MB	0.06s / 374MB	0.07s / 448MB	0.07s / 581MB	0.11s / 917MB
members	0.01s / 19MB	0.03s / 55MB	0.05s / 118MB	0.11s / 174MB	0.11s / 180MB	0.14s / 218MB	0.21s / 285MB
hierarchy	0.01s / 20MB	0.03s / 55MB	0.05s / 127MB	0.10s / 174MB	0.11s / 177MB	0.15s / 194MB	0.21s / 294MB
disasm	0.01s / 19MB	0.03s / 58MB	0.05s / 124MB	0.10s / 175MB	0.11s / 182MB	0.14s / 193MB	0.20s / 281MB
getclass	0.02s / 33MB	0.21s / 131MB	0.18s / 256MB	0.41s / 515MB	0.50s / 527MB	0.70s / 830MB	1.30s / 1200MB

Documentation

Architecture	crates, the lift/structure/emit pipeline, DEX versions, invoke-custom (中文)
CLI reference	every option and subcommand, in detail (中文)
Subcommands	the progressive-analysis reference (中文)
Benchmarks	timings, memory, methodology, ASC comparison (中文)
Performance engineering	5min → 6s: six rounds, measured dead ends (中文)
Corpus validation	39 real APKs, 4.08M files, javac parse-clean, bugs found & fixed (中文)

Known limitations

Erased types (DEX has no Signature); d8-desugared -$$Lambda$ classes stay separate files; a few R8 monster methods degrade via timeout (and a handful of deadline-truncated ones can vary between runs); pattern-switch renders as desugared dispatch chains. The javac gate above is a SYNTAX gate — semantic diagnostics (missing Android classpath, and in rare register-heavy monster methods a type-confused local) remain in the output. For scale: a full-classpath javac battery over four corpora currently sits at ~7,500 ddc-attributable errors across 0.91M files (v0.1.19 measured 11,345 — see benchmarks); jadx on the same battery lands at roughly half that, the gap concentrated in type recovery. Details: architecture.

Testing

cargo test    # 74 tests

License

MIT © ejfkdev