What does REA setup install?
REA, or Reverse Engineer Anything, has three layers: the skill teaches an agent the workflow, the MCP registration tells a client how to start REA, and analysis engines provide capabilities such as decompilation. A successful setup does not mean every engine is installed or that every model gateway accepts the tool schemas.
For a first installation, select one client explicitly. This guide checks REA 6.7.0 with Claude Code configuration on macOS. It does not claim that Claude Desktop, Codex, Windows and Linux have all been tested by WebFV.
Check the runtime before editing configuration
Run node --version. In our test, Node 24.5.0 caused setup to return needs_human without planned or applied changes. The remediation required Node 22.x at least 22.19, Node 24.x at least 24.11, or Node 26+. Repeating the same command with Node 24.11.1 produced a valid plan.
This matters because the output can include missing-engine checks even when the immediate blocker is the runtime. Read the top-level remediation before installing unrelated software. Use a supported Node release and confirm the same runtime is available to the client process, not just an interactive terminal.
Preview a bounded installation
A separate Claude profile keeps this first check away from your existing configuration. Replace the example path with a private directory you control. Use the same profile path for setup, diagnosis and the Claude session.
CLAUDE_CONFIG_DIR="$PWD/claude-rea-profile" \
npx -y [email protected] setup \
--client claude_code --dry-run --json
Read the plan. For the isolated profile in our test, it proposed a REA MCP registration in claude-rea-profile/.claude.json and a skill under claude-rea-profile/skills/reverse-engineer-anything. No analysis-engine installation was requested. Confirm every target path before applying it.
CLAUDE_CONFIG_DIR="$PWD/claude-rea-profile" \
npx -y [email protected] setup \
--client claude_code --yes --json
The --yes flag approves the planned setup actions. Do not substitute --all-detected unless you intend to modify every detected client. If you normally use the default Claude profile, leaving CLAUDE_CONFIG_DIR unset is a deliberate choice to use that profile instead.
Inspect the generated registration
Our isolated setup wrote this server definition:
{
"mcpServers": {
"rea": {
"command": "npx",
"args": ["-y", "[email protected]", "mcp"]
}
}
}
The version is pinned in the command. A separately updated global REA installation does not automatically change this registration. Check the command that the client actually starts when investigating a version mismatch.
You can check a sanitized copy in the MCP config checker. A valid JSON result establishes structure only. It cannot find your Node executable, launch REA, authenticate a model provider or prove that tools are visible inside Claude.
Verify in the same profile
CLAUDE_CONFIG_DIR="$PWD/claude-rea-profile" \
npx -y [email protected] doctor \
--client claude_code --skill --json
CLAUDE_CONFIG_DIR="$PWD/claude-rea-profile" claude mcp list
The first command checks selected readiness and installed workflow files. The second asks Claude Code for actual MCP connection status. Start or restart Claude with that same profile after changing registration. Keep the outputs separate: setup completion, doctor readiness, client connection, and a successful analysis are different results.
Our tested setup reached ready and created the expected registration and skill files. A separate stdio MCP client initialized the server as version 6.7.0 and received 207 tools from tools/list. This checks the server transport and tool catalog, not the Claude interface or a paid model request. That is a configuration result, not proof of an authenticated model session or a working Ghidra/IDA/Hopper workflow.
When setup still fails
If the plan targets the wrong directory, stop and fix the profile selection. If Node is unsupported, fix the runtime first. If a client cannot start the configured command, compare its environment and PATH with your terminal. If tools connect but an analysis is unavailable, inspect the selected provider checks in REA doctor.
For schema recursion or maximum-size errors from a model gateway, use the separate REA schema error guide. Reinstalling every engine does not address that class of failure.
Sources: official installation guide, REA getting started, and 6.7.0 release. Instructions are version-specific; review a new release's plan before updating an existing profile.