Docker Reverse Engineer Anything MCP Server

docker-reverse-engineer-anything

REA (Reverse Engineer Anything) in a Docker image, with everything it needs to tear apart Android APKs already inside, and a Streamable HTTP MCP endpoint in front of it.

REA is morluto's MCP server and CLI that lets an AI agent inspect software it has no source for: decompile APKs, trace references, read resources, poke at devices over adb, dig through Electron bundles, .NET assemblies, Go and Rust binaries. Upstream ships it as the rea-agents npm package that talks MCP over stdio and expects you to install a JDK, JADX and Apktool yourself. This image does that part for you.

This is an unofficial packaging. REA itself lives at morluto/rea and is MIT-licensed. Bugs in REA's tools go there; bugs in the image or the HTTP bridge go here.

Table of Contents

• What's in the image
• Quick start
• Ways to run it
• Opt-in tools
• Configuration
• Security
• Agent integrations
• Versions and updates
• Development
• License

What's in the image

APK analysis needs no extra setup: inspect_android_package, search_android_classes, inspect_android_class, inspect_android_method, trace_android_references and decode_android_resources work out of the box, and so do REA's static JavaScript/Electron, .NET, Go, Rust, WASM, SQLite and capture tools.

Native-binary decompilation (the .so libraries inside an APK, or any other ELF, PE or Mach-O) needs Ghidra. It is free but adds about 900 MB, so it is not baked in: set REA_INSTALL=ghidra and the container fetches it on first start. See Opt-in tools.

Images are built for linux/amd64 and linux/arm64. CI runs the APK tests on both architectures. Ghidra is amd64 only.

Quick start

export REA_BRIDGE_TOKEN="$(openssl rand -hex 32)"
mkdir -p targets && cp /path/to/app.apk targets/

docker run -d --name rea \
  -e REA_BRIDGE_TOKEN \
  -p 127.0.0.1:8080:8080 \
  --read-only \
  --tmpfs /tmp:rw,nosuid,size=4g \
  --tmpfs /home/rea:rw,nosuid,size=256m \
  --cap-drop ALL \
  --security-opt no-new-privileges \
  -v "$PWD/targets:/work:ro" \
  psyb0t/reverse-engineer-anything:latest

claude mcp add --transport http rea http://localhost:8080/mcp \
  --header "Authorization: Bearer $REA_BRIDGE_TOKEN"

Then ask your agent something like "decompile the login flow in /work/app.apk and show me where the token gets stored". Paths are container paths: whatever you mount at /work is what REA sees.

Ways to run it

The entrypoint takes one argument:

Streamable HTTP bridge

The bridge is a small Node service in front of rea mcp. Each MCP session gets its own REA process, because REA keeps analysis state per process. Messages pass through untouched, progress notifications included. Sessions end on DELETE, after the idle timeout, or when the container stops, and their REA process goes with them.

stdio MCP

For clients that only launch local commands:

{
  "mcpServers": {
    "rea": {
      "command": "docker",
      "args": [
        "run",
        "-i",
        "--rm",
        "--read-only",
        "--tmpfs",
        "/tmp:rw,nosuid,size=4g",
        "--tmpfs",
        "/home/rea:rw,nosuid,size=256m",
        "--cap-drop",
        "ALL",
        "--security-opt",
        "no-new-privileges",
        "-v",
        "/absolute/path/to/targets:/work:ro",
        "psyb0t/reverse-engineer-anything:latest",
        "mcp"
      ]
    }
  }
}

CLI

docker run --rm --network none --read-only \
  --tmpfs /tmp:rw,nosuid,size=4g --tmpfs /home/rea:rw,nosuid,size=256m \
  --cap-drop ALL --security-opt no-new-privileges \
  -v "$PWD/targets:/work:ro" \
  psyb0t/reverse-engineer-anything:latest \
  search-android-classes /work/app.apk MainActivity

Output is JSON. --help lists every command.

Opt-in tools

Some tools REA can drive are free to redistribute but too big to put in every image. Name them in REA_INSTALL and mount a volume at /opt/rea-tools:

docker run -d --name rea \
  -e REA_BRIDGE_TOKEN \
  -e REA_INSTALL=ghidra \
  -v rea-tools:/opt/rea-tools \
  -p 127.0.0.1:8080:8080 \
  --read-only \
  --tmpfs /tmp:rw,nosuid,size=4g \
  --tmpfs /home/rea:rw,nosuid,size=256m \
  --cap-drop ALL \
  --security-opt no-new-privileges \
  -v "$PWD/targets:/work:ro" \
  psyb0t/reverse-engineer-anything:latest

On start, before REA runs, the entrypoint installs each listed tool that is not already in the volume. It downloads the version pinned in the image, refuses it unless the SHA-256 matches, unpacks it, and points REA at it. Later starts find it in the volume and skip the download. It works in every mode, mcp and the CLI included, and logs to stderr only.

The first start needs network access to GitHub. Without a volume, a --read-only container has nowhere to put the tool and exits with a message saying so; a writable one installs into its own layer and downloads again for every new container. Ghidra's official release has no arm64 decompiler, so on arm64 REA_INSTALL=ghidra refuses with that reason instead of half-working.

Bring your own

Paid or login-gated tools are never downloaded. Install them yourself and point REA at them:

• Ghidra builds of your own (an arm64 build, a newer version): mount it and set GHIDRA_INSTALL_DIR. A set GHIDRA_INSTALL_DIR wins over REA_INSTALL=ghidra.
• IDA: mount your installation and set REA_IDA_MCP_CONFIG, per REA's IDA guide.
• JEB: run its MCP server where REA can reach it on loopback (a sidecar sharing the container's network namespace) and set REA_JEB_MCP_URL.
• Hopper is a desktop app and is not supported here.

Configuration

REA's own variables pass straight through, for example REA_JADX_HEAP_MIB and REA_JADX_OPERATION_TIMEOUT_MS for big APKs, REA_LOG_LEVEL, or REA_MCP_MAX_RESPONSE_BYTES. See REA's documentation. The image presets JAVA_HOME, REA_JADX_MCP_JAR, REA_APKTOOL_COMMAND and REA_ADB_PATH.

GET /healthz answers without a token: {"status":"ok","sessions":N}.

Security

REA can run, drive and instrument the software it looks at, and it does no permission checks of its own. The container is the boundary, so keep it tight:

• The bridge won't start without a token, compares it in constant time, checks the Host header, and never hands the token to REA.
• The image runs as an unprivileged rea user. Run it the way the examples do: read-only root, tmpfs for /tmp and /home/rea, all capabilities dropped, no-new-privileges, targets mounted read-only, port bound to 127.0.0.1.
• /tmp can't be noexec: the JVM loads native libraries from it.
• Exposing it beyond localhost? Put TLS in front and add the hostname to REA_BRIDGE_ALLOWED_HOSTS.
• Only analyze software you own or are authorized to examine. REA's disclaimer applies.

Agent integrations

• Skill for Claude Code, Codex and OpenClaw: claude plugin install reverse-engineer-anything@psyb0t, codex plugin add reverse-engineer-anything@psyb0t, or from ClawHub. It teaches the agent the APK workflow and ships a locked-down CLI wrapper. Add the marketplace first with claude plugin marketplace add psyb0t/agents.
• OpenClaw plugin @psyb0t/reverse-engineer-anything: a stdio bridge to a running container. Set REA_URL and REA_BRIDGE_TOKEN.
• MCP registry: io.github.psyb0t/reverse-engineer-anything.

Versions and updates

The image version is this repository's own (VERSION); REA's version is in the table above and in package.json. REA ships releases almost daily. This image only takes a REA release, or any other npm package, once it is at least seven days old, installs with lifecycle scripts disabled from a frozen lockfile, and checks every downloaded jar against a pinned SHA-256. The opt-in tools follow the same rule: one version per image release, at least seven days old, checked against its SHA-256. Base images are pinned by digest.

Development

The host needs Docker, Make, Git and a shell. Everything else runs in containers.

make help        # every target
make lint        # tsc, eslint, prettier, shellcheck, hadolint
make test        # bridge unit tests + the image against a real APK
make build       # build psyb0t/reverse-engineer-anything locally
make run         # run the bridge on 127.0.0.1:8080 (needs REA_BRIDGE_TOKEN)
make pkg-update PKG=rea-agents   # move to the newest REA that clears the age gate

The integration test starts the hardened container, checks that a missing token gets a 401, then drives REA's Android tools through the bridge against Appium's ApiDemos APK. It also starts a container with REA_INSTALL=ghidra and has Ghidra analyze a binary through the bridge (on arm64 it checks for the refusal instead). The Ghidra download is cached in the rea-integration-tools volume; make clean removes it.

License

WTFPL for this repository's code. REA is MIT, and the bundled tools keep their own licenses: see THIRD_PARTY_NOTICES.md.