Skip to content

t1k:rea

FieldValue
Modulet1k-reverse-engineering
Version2.17.24
Effortmedium
Tools—

Keywords: apk, asar, assetripper, binary analysis, decompile, disassemble, electron, firmware, ghidra, global-metadata, har, hopper, ida, il2cpp, ipa, jadx, libil2cpp, malware, pcap, rea, rea-agents, reverse engineer, reverse engineering, unity, xapk

/t1k:rea
<target path|url> [question]

REA — Reverse Engineer Anything, With Evidence

Section titled “REA — Reverse Engineer Anything, With Evidence”

REA (morluto/rea, npm rea-agents, MIT, site rea.tools) is an MCP server and CLI that turns a local artifact into citable evidence: symbols, callers, strings, resources, wire traffic. Analysis runs locally on the file; nothing is uploaded.

Pin: rea-agents@6.7.0, audited 2026-10-20 — no install hooks, SLSA provenance present, no telemetry strings, network egress only to GitHub and hopperapp.com. This is a snapshot of a tool that ships weekly: re-audit after any major bump before trusting the pin.

IntentGo to
Install, register or upgrade REASafe setup
Which command for this targetreferences/target-map.md
Unity IL2CPP build (.apk/.ipa, libil2cpp.so)references/unity-il2cpp.md
Understand what a feature doesTrace a feature
“Port this app’s code/assets to ours”Legal limits, then decline
Port a game the studio holds the rights to (partner permission on record)t1k:apk-port
Confused about a flag or the output shaperea --llms, rea <cmd> --schema

REA’s own DISCLAIMER excludes: DRM/licence/anti-cheat circumvention, game cheats, copying proprietary software or reproducing its protected code and assets, and auth/access-control bypass on systems you are not authorised to test. That exclusion is the skill’s boundary, not a suggestion.

  • Use REA on software the user owns or is authorised to examine, on malware and incident samples, on CTFs, and to answer interoperability questions (file formats, protocols, schemas).
  • A request to decompile a third-party commercial app and port its code or assets is a decline. Say so in one line, then offer the clean-room route: write a behavioural spec from public docs and observed behaviour, and build original code, names and assets against that spec (t1k:cleanroom is the studio’s clean-room workflow).
  • A game whose rights holder has already granted the studio permission is not a third-party port: route it to t1k:apk-port, which records the grant and keeps SDK ids and the publisher’s identity out of the port.
  • Passive observation of a public website is fine. Never log in, bypass bot protection, or replay another party’s session tokens.
  • Ownership claims are not on record unless the user states them. Do not infer permission from “it’s on my machine” or from the file being present.

When a request sits on the boundary, name the boundary and ask which side it is on before running anything. An artifact inspected for the wrong reason is not undone by a later apology.

NEVER run bunx rea-agents setup blind: it edits agent configs. Dry-run first, in an isolated HOME, and read the plan before letting it touch anything real.

  1. Dry-run in a throwaway HOME (scratch via the kit helper, never a bare temp dir):

    Terminal window
    B=$(bun .claude/scripts/t1k-scratch.cjs new --purpose rea-setup)
    mkdir -p "$B/home" && printf '{}\n' > "$B/home/.claude.json"

HOME=“$B/home” bunx —ignore-scripts rea-agents@6.7.0 setup —all-detected —yes —dry-run

The verified plan for Claude Code (client id `claude_code`) is exactly two things: add one MCP
server entry to `~/.claude.json` (existing config preserved, backup written as `.rea.backup`),
and install REA's own skill `reverse-engineer-anything` into `~/.claude/skills/`. Anything else
in the plan is a surprise: stop and report it instead of confirming.
2. Prefer project-scoped registration, which changes no global config:
`rea mcp add --agent claude-code --no-global`, then `rea doctor` to prove the provider resolves.
3. Pin `rea-agents@6.7.0` in `.mcp.json` so a run is reproducible; `rea update` upgrades with
verification rather than silently floating.
4. Never use the curl-pipe `install.sh` path. Always `--ignore-scripts` when only inspecting the
package, and extract a tarball into its own empty directory.
Requirements to check before blaming the tool: Node 22.19+ / 24.11+ / 26+; a procps `ps` on PATH
on Linux (JS, Ghidra and JADX use it for ownership checks); Hopper, Ghidra or IDA already installed
for native work (`GHIDRA_INSTALL_DIR`, `JAVA_HOME`); headless JADX plus a full JDK for APKs. REA
never installs Ghidra or Java, so a missing backend is a missing prerequisite, not a bug.
Two traps in the setup flags: `rea setup --client <id>` needs the exact integration id
(`claude_code`, not `claude`) and answers `Unsupported agent integration` otherwise, while
`rea mcp add --agent claude-code` uses the dashed form. First Hopper run shows a licence dialog
that must be completed in the same user session, and a large binary can need
`REA_GHIDRA_STARTUP_TIMEOUT_MS` / `REA_HOPPER_STARTUP_TIMEOUT_MS` raised.
## Which command
Full matrix, per-target notes and the static-vs-runtime split: `references/target-map.md`. Start
here:
| Target | First command |
|---|---|
| Unknown artifact or app directory | `rea analyze <path>` |
| Electron / JS app, ASAR bundle | `rea analyze-javascript-application <path> --json` |
| Native binary (Mach-O / ELF / PE) | `rea inspect <path>` |
| Android APK | `rea inspect-android-package <path>` |
| .NET assembly, Unity IL2CPP metadata | `rea inspect-managed-artifact` |
| Website, HAR / SAZ / pcap | `rea inspect-web-network-capture`, `rea inspect-pcap-capture` |
Static commands read files only. `capture-process`, `instrument-with-frida` and
`observe-native-calls` EXECUTE the target with the user's permissions, and REA is not a sandbox:
run untrusted samples in a VM, never on the workstation.
## Trace a feature
1. `rea doctor` and `rea providers` — confirm a backend exists for the artifact class before
spending minutes on a doomed run.
2. Inventory first (`analyze` / `inspect-artifact`), then narrow with `search` or `trace` on a
literal string, class name or route. Guessing at symbols before inventory wastes the run.
3. Read the code (`function`, `decompile`, `instructions`) and quote what came back. Report the
limitations REA states about the result; a decompiler's failure is evidence too.
4. Export Evidence bundles with `rea evidence-export` so a later step can diff with `rea compare`.
5. Write findings as a spec — behaviour, data format, protocol shape — never as copied code.
## Executed by
Setup, inventory and evidence export are mechanical command runs; judging what the evidence means,
and deciding whether a request is inside the legal boundary, is not. Routing:
`docs/orchestration-rules.md` § "Task-Type → Agent Routing".
| Step | Agent |
|---|---|
| Run the setup dry-run, `rea doctor`, inventory and `evidence-export`, relay the JSON | `t1k-data-verifier` |
| Decide the legal boundary, read decompiled code, write the spec | the invoking session (reasoning tier, not cheap) |
## References
- `references/target-map.md` — every target class, its commands, and the per-class caveats.
- `references/unity-il2cpp.md` — Unity IL2CPP end to end: metadata-version pinning, what the dummy
DLLs do and do not give you, AssetRipper over its HTTP API, and the three silent Ghidra traps
(image base, body clamping, invalid entries) that make pseudocode look real when it is not.