REA: Reverse Engineer Anything with Agents
Hands reverse engineering to your coding agent, from app behavior down to native binaries.
- Source repo
- morluto/rea
- Stars
- ★ 471k
- Last updated
- today
- License
- MIT
- Primary language
- TypeScript
- FA score
- 84/100 · Good
Read the setup guide: MCP connection, first workflows and troubleshooting →
At a glance
- How it runs
- Works with
- Universal · cross-platformCodex · Claude Code · Claude.ai
- Cost
- Free, no paid service needed
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- A product team sees a search or offline-caching feature in a shipped app with no source access and wants the agent to explain it, show evidence, and rebuild it for their own stack.
- Not a fit if
- Developers who only read source code and never touch binaries or runtime behavior
- Teams that need deep native binary analysis on Windows (only a limited P0 boundary exists)
- Users unwilling to install Hopper locally or supply their own Ghidra/JDK
- Website
- rea.tools ↗ set on GitHub
- Maintenance
- Commits in the last 30 dayslast commit 2026-11-20latest release rea-agents-6.10.1 (2026-11-20)+1998.4 ★/day over 7 days
- Source review
- 84/100 · Good
What does this agent do, and when should you use it?
REA is a reverse-engineering toolkit for coding agents, shipped as the npm package rea-agents with both a rea CLI and an MCP server that share one evidence contract and one investigation model. It wraps a deep-analysis provider — Hopper on macOS/Linux (demo mode supported) or a bring-your-own Ghidra 12.1.4 on Linux/macOS plus an experimental Windows x64 P0 — and exposes it through tools such as open_binary, binary_overview, search_strings, find_xrefs_to_name, get_call_graph, procedure_pseudo_code, and batch_decompile. Beyond native analysis it covers passive CDP website observation, attach-only Node/Electron V8 Inspector observation, JavaScript and source-map reconstruction, ASAR/package structure mapping, managed PE/CLI triage, and controlled process capture with comparison. Every successful call is recorded as deterministic Evidence with artifact and provider identity, confidence, authority, and limitations, exportable as bundles or reusable as snapshots. Analysis runs entirely on your local host — REA provides no hosted service and never uploads the target. It deliberately does not claim to recover original source code.
After setup, REA probes for local agents (Claude Code, Claude Desktop, Codex, Cursor, Gemini CLI, Windsurf, Devin), writes MCP registrations additively and backup-first, and installs the matching routing skill. On each deep-analysis open it resolves a provider — request-level --provider/provider_id beats REA_ANALYSIS_PROVIDER, auto binds only a sole usable candidate and reports ambiguous otherwise, and runtime failure never triggers silent fallback. Commands like rea analyze, rea decompile, rea function, rea xrefs, and rea trace return the same Evidence envelopes as MCP; an MCP session can retain a live target and evidence ledger, and snapshot_path atomically imports a snapshot in open_binary or saves one in close_binary. Other surfaces include inspect-managed-artifact for PE/CLI metadata and CIL hashes, list-browser-targets/inspect-web-page for passive CDP capture, list-javascript-runtime-targets/observe-javascript-runtime for lifecycle metadata only, and capture-process/capture-browser-scenario for explicitly approved, per-call-granted behavioral capture.
- A product team sees a search or offline-caching feature in a shipped app with no source access and wants the agent to explain it, show evidence, and rebuild it for their own stack.
- A reverse engineer or security analyst needs to trace a suspicious string, symbol, or IPC channel all the way to the routine that implements it while keeping an auditable evidence chain.
- A developer handed an Electron app wants to statically reconstruct its ASAR modules, entrypoints, preload, and IPC graph without executing bootstrap code.
- Someone comparing two releases needs to find the first behavioral divergence across terminal output, filesystem, or protocol dimensions rather than eyeballing diffs.
- A maintainer of legacy .NET/CLI assemblies wants read-only metadata and P/Invoke boundary triage without loading the assembly or resolving CLR dependencies.
- A team that must reproduce analysis wants to freeze successful calls into snapshots or Evidence bundles and replay the same query without relaunching a provider.
Key setup facts
Pulled from the install and usage notes below. The project README has the current values.
- Install / run
npx --yes rea-agents@latest setupnpm install --global rea-agentsnpx -y rea-agents@latest analyze /Applications/Notes.appnpx -y rea-agents@latest inspect /Applications/Notes.app
- Environment variables
GHIDRA_INSTALL_DIRTARGET_ID- Runtime
- Node.js 22.19+
- Install notes cover
- Windows · macOS · Linux
How do you install or deploy this agent?
The recommended path is the reviewed setup wizard, which prints its full plan, exact paths, and external effects before a final confirmation that defaults to No:
npx --yes rea-agents@latest setupTo get a shell-visible rea command instead:
npm install --global rea-agents
rea setup
rea doctorSetup configures Claude Code, Claude Desktop, Codex, Cursor, Gemini CLI, and Windsurf when detected; Devin is reported but left unchanged. Deep native analysis needs a provider: setup can install Hopper (official DMG to ~/Applications on macOS; pinned 6.4.2 package with checksum verification on Ubuntu 24.04+, Fedora 41+, Arch), while Ghidra is strictly bring-your-own:
export GHIDRA_INSTALL_DIR=/absolute/path/to/ghidra_12.1.4_PUBLIC
export JAVA_HOME=/absolute/path/to/jdk-21 # optional when java and javac resolve from PATH
rea doctor --json
rea providers --jsonRequirements: macOS 12+, Ubuntu 24.04+/Fedora 41+/64-bit Arch Linux, with Windows x64 only for the experimental Ghidra P0; Node.js 22.19+ or 24.11+.
How do you use this agent?
After setup, restart any configured agent so it loads the aligned MCP registration and routing skill, then describe what you want investigated:
Reverse engineer the Notes app. Find how offline search works, explain it,
and build a version for my project using TypeScript and SQLite.For a one-off terminal overview:
npx -y rea-agents@latest analyze /Applications/Notes.app
npx -y rea-agents@latest inspect /Applications/Notes.app
npx -y rea-agents@latest trace /Applications/Notes.app "offline"Manual configuration for any client that supports local MCP servers:
{
"mcpServers": {
"rea": {
"command": "npx",
"args": ["-y", "[email protected]", "mcp"]
}
}
}Reuse an approved result without relaunching a provider:
rea analyze /absolute/path/to/app --snapshot /absolute/path/to/analysis/app.jsonBrowser and runtime observation are off by default and need literal loopback endpoints plus approved origins:
export REA_BROWSER_OBSERVE_ENABLED=true
export REA_BROWSER_CDP_ENDPOINTS_JSON='["http://127.0.0.1:9222"]'
export REA_BROWSER_ALLOWED_ORIGINS_JSON='["http://127.0.0.1:3000"]'
rea inspect-web-page http://127.0.0.1:9222 TARGET_ID --approved --jsonWhat are this agent's strengths and limitations?
- One workflow and one evidence contract span the CLI and the MCP server, so the same capability can be a single terminal call or a session-held target with a live evidence ledger.
- Providers are optional and explicitly bound: bring your own Ghidra 12.1.4 for a read-only path, or let setup install Hopper; switching requires an explicit selector and runtime failure never silently falls back.
- Analysis stays local with no hosted service, and every dynamic surface — process capture, browser scenarios, JS runtime observation — is disabled by default and gated by per-call approval.
- Coverage extends well past a single disassembler: passive CDP web observation, attach-only Node/Electron V8 Inspector observation, JavaScript/source-map reconstruction, and read-only managed PE/CLI triage.
- Each successful call becomes deterministic Evidence carrying artifact and provider identity, confidence, and limitations, and can be exported, compared, or replayed from an immutable snapshot.
- Deep native analysis depends on external tooling: Hopper is separate licensed software (demo mode is vendor-limited), and Ghidra must be supplied at exactly 12.1.4 with a 64-bit full JDK 21 or doctor rejects it.
- The platform boundary is real: full capability centers on macOS and Linux, while Windows offers only an experimental x64 Ghidra P0 that does not claim named-pipe DACL or hostile-local-user isolation.
- The runtime is heavyweight — Node.js 22.19+ or 24.11+, and Linux demo automation additionally needs Xvfb, Python 3, libX11.so.6, and libXtst.so.6 — so first-time preparation is not trivial.
- REA does not recover original source; decompiler text is explicitly a provider-specific observation, and treating it as source will mislead you.
- Several advanced capabilities (browser scenarios, process capture, JS runtime observation, snapshots) require extra environment variables and per-call grants, so you pay an upfront cost in the permission model.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| REA: Reverse Engineer Anything with Agents This agent | 84 · Good | CLIFree | ★ 471k | today | TypeScript | Codex · Claude Code · Claude.ai |
| REA Reverse Engineering Workbench | 55 · Major gaps | CLIFree | ★ 103 | 1mo ago | HTML | Codex · Claude Code |
| Auto RE Agent | 70 · Some gaps | CLIFree + model costs | ★ 2.5k | 2mo ago | Python | Codex · Claude Code · OpenAI API · Claude API |
| Eleven Reverse Engineering Workbench | 63 · Some gaps | Desktop appFree | ★ 1k | 2d ago | TypeScript | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
AGENTS.md describes REA as a local tool and directs work to respect ownership and capability boundaries. The README and installation guide say runtime capture uses caller permissions and that optional Hopper installation and configuration writes require separate approval, supporting strong least-privilege and confirmation scores. The docs explain local files, network origins, sensitive-value handling, and rollback limits. The packet contains no dependency-audit evidence, and runtime capture and environment export still have external effects, so those scores are not full. Plugin metadata names an author, repository, and version, but publisher identity is unverified and cannot be treated as confirmed provenance.
The README, installation guide, and source agree on MCP configuration, schema-profile settings, and startup error handling; the guide describes runtime dependencies and recovery paths. Startup exceptions are hidden behind a generic message, limiting direct diagnosis even as it reduces information exposure. Many external tools are caller-supplied, and the packet cannot establish their availability.
The README covers static binaries, applications, websites, devices, and runtime analysis, with tools and host requirements for each. The packet distinguishes static inspection from workflows that run or interact with targets, and states Node, operating-system, and optional-provider constraints. Skill descriptions are specific, but the provided skills are a sample and the core skill is not included, so trigger coverage cannot receive full marks.
The repository has multilingual entry points, domain-organized guides, installation and troubleshooting instructions, examples/FAQ, and stated limitations. The license is MIT. The plugin manifest and README show version information, but this packet does not provide enough changelog and maintenance-responsibility evidence for full scores.
The product exposes a broad set of reverse-engineering capabilities through CLI/MCP, returning evidence and limitations that agents can use to explain or reproduce behavior. The range of external engines and tools increases setup and maintenance costs. Static source material cannot establish the quality of actual analysis results.
The README, guides, AGENTS.md, plugin manifests, and source mutually support claims about product scope, setup approval, boundaries, and error handling. The docs distinguish observations, derivations, inferences, unknowns, and evidence sources. The packet includes concrete verification notes and limits, but a static review cannot independently reproduce those records, so confidence remains low.
- Runtime capture may run or interact with a target using the current user's permissions; confirm the target and test data are authorized for that use.
- Many deep-analysis workflows depend on caller-installed or caller-started tools; check each guide for host, version, and capability limits.
- Publisher identity has not been verified by the FollowAgents enterprise registry; repository metadata alone does not establish it.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
rea-agents-6.10.1 rea-agents: v6.10.1
Release notes excerpt
[6.10.1](https://github.com/morluto/rea/compare/rea-agents-6.10.0...rea-agents-6.10.1) (2026-11-20) This is the first 6.10 release published to npm. It includes the changes, migration notes, and contributor credits from 6.10.0 below, plus the patch fix. Bug Fixes * kuna: use one constant for the serial job count ([#12735](https://github.com/morluto/rea/issues/12735)) ([3e3faf2](https://github.com/morluto/rea/commit/3e3faf2fc32dbadc96a45f51f4361eb64ce6ed32)) --- [6.10.0](https://github.com/morluto/rea/compare/rea-agents-6.9.0...rea-agents-6.10.0) (2026-11-20) Highlights …View on GitHub ↗ -
rea-agents-6.10.0 rea-agents: v6.10.0
Release notes excerpt
[6.10.0](https://github.com/morluto/rea/compare/rea-agents-6.9.0...rea-agents-6.10.0) (2026-11-20) Highlights - Inspect and extract UEFI firmware volumes, files, and AMI NVRAM with three offline CLI and MCP tools. [#12342](https://github.com/morluto/rea/pull/12342). Thanks @bartman081523. - Inspect Android boot images and OTA payloads, NES 2.0 containers, and Maven, Gradle, and NuGet package metadata. [a88c0a5](https://github.com/morluto/rea/commit/a88c0a55544e16a710ddaad2feb5a2c6e7011f4e) [#9487](https://github.com/morluto/rea/pull/9487) [#11378](https://github.com/morluto/rea/pull/1 …View on GitHub ↗ -
rea-agents-6.9.0 rea-agents: v6.9.0
Release notes excerpt
[6.9.0](https://github.com/morluto/rea/compare/rea-agents-6.8.0...rea-agents-6.9.0) (2026-11-11) Highlights - Build and test desktop tasks from recorded application state, including downloaded Linux apps, reusable build computers and multiple tasks per capture. - Inspect more shipped artifacts offline: standalone application bundles, embedded firmware, NTFS acquisitions, API contracts and V8 code-cache containers. - Keep native and runtime investigations grounded in retained Evidence, with paged results, explicit coverage and recoverable resource limits. Changes …View on GitHub ↗
FAQ
Is REA free? What about Hopper?
Does it upload my application anywhere?
Can REA recover the original source code?
Does it work on Windows?
rea setup plus Hopper installation remain unavailable on Windows.Can I reproduce and diff analysis results?
rea evidence-import/rea evidence-export, then compared with rea compare. You can also freeze calls into provider-neutral snapshots (--snapshot, or snapshot_path in MCP) that are reused only when the binary digest, kind, format, architecture, operation parameters, and provider build all match.Is REA: Reverse Engineer Anything with Agents open source?
Is REA: Reverse Engineer Anything with Agents safe to use?
Is REA: Reverse Engineer Anything with Agents still maintained?
Who makes REA: Reverse Engineer Anything with Agents?
What is REA: Reverse Engineer Anything with Agents's official website?
Thanks — we use this to improve the page.