VarwoREA MCP connects an AI agent to REA's software-inspection tools. Start by selecting one client and checking its registration. Seeing a server in the client is a useful first check; native analysis may still need a separate engine.
Official setup reference · REA project overview · Claude Code walkthrough
Documentation reviewed October 23, 2026. These are reviewed setup examples, not an installation or live agent connection performed by Varwo.
Choose a client before configuring MCP
REA's setup supports explicit client selectors. Use the selector for the application that will actually run your investigation, rather than configuring every detected agent. The upstream installation guide documents these choices:
| Your client | Setup selector | Follow-up |
|---|---|---|
| Codex | codex | Check the selected profile and restart the intended session |
| Cursor | cursor | Check the registration in the Cursor configuration you actually use |
| Claude Code | claude_code | Follow the dedicated Claude guide |
The client provides the agent environment. REA provides inspection tools. Neither MCP registration nor a package download creates access to a paid model or an analysis-engine license.
Preview setup for one client
With an appropriate Node/npm environment available, this example asks for a Cursor-specific plan:
npx rea-agents@latest setup --client cursor --dry-runRead the proposed paths and operations. The package runner may download the package even though the setup itself is a dry run. To select Codex, substitute codex for cursor.
When the plan matches your intended scope, the corresponding apply command is:
npx rea-agents@latest setup --client cursorReview and approve the setup's actual changes, then restart the named client. Setup is client-wide by default; using a repository's terminal does not establish project-only installation. For profile paths and scoped alternatives, use the client configuration reference.
Check registration and analysis readiness separately
Use the same client choice for the diagnostic:
npx rea-agents@latest doctor --client cursor --jsonCheck the result in the client/profile where you applied setup. If you also need to check the installed workflow, consult the installation guide's doctor --skill scope. Do not treat an unrelated profile's result as a check of the active session.
For a native binary, the selected provider and its prerequisites are another check. Static JavaScript/Electron inspection can use a direct CLI route without an agent registration; the CLI reference explains that path.
Keep transport and provider choices distinct
REA's MCP contracts describe stdio as the normal registration transport. The optional HTTP mode is a different, explicitly enabled local transport; it is not a hosted public REA service to paste into every client's URL field.
Start with the documented registration rather than inventing an HTTP endpoint. Likewise, choose an analysis engine for the target artifact instead of assuming that the MCP transport performs decompilation itself.
A first task with an evidence check
Use a small application you own or are authorized to inspect. Give the agent one question:
Identify the exact build of this local demo application.
Investigate its CSV export for a comma and a line break in one field.
Return the relevant evidence locations, your proposed export rule,
and a tiny input/output check. Mark unknowns. Do not modify the app.Compare the explanation with an export from that same build. Keep artifact identity, observed behavior and interpretation separate. A fluent explanation without supporting evidence is an incomplete result.
Troubleshoot the layer that failed
| What you observe | Check next |
|---|---|
| No REA server in the agent | Active client/profile, setup plan and that client's registration diagnostic |
| Server visible, analysis cannot start | Actual target type, selected provider and its prerequisites |
| Schema or tool-list rejection | Current client guidance and the MCP contracts; preserve unrelated settings |
| A result with missing coverage | The named limitation and the missing artifact or observation |
Do not repair a connection by indiscriminately changing every MCP entry. Save the failing command, relevant diagnostic and expected scope so the next check tests a concrete explanation.
Frequently asked questions
Do I need both this page and the Claude Code guide?
Use this page to choose a client and understand the setup layers. The Claude Code guide covers that client's particular workflow and checks.
Does a successful connection mean REA can recover any application's source?
No. Tool availability, artifact coverage and the evidence returned determine what an investigation can establish. See the REA overview for a bounded starting task.
How does REA compare with a hosted binary-analysis service?
See the AI reverse-engineering tool comparison for REA and RevEng.AI's different integration routes.