elenchus

Elenchus is an MCP server that orchestrates an adversarial debate between Verifier and Critic agents to verify code and identify issues. It uses a structured Verifier→Critic loop to produce evidence-backed findings across multiple rounds, with optional LLM-based convergence and severity evaluation. The server exposes a set of MCP tools for session management, context retrieval, and structured verification tasks. To use it, add the Elenchus MCP server to your MCP client configuration (for example in Claude Desktop, Claude Code, VS Code, or Cursor) so your assistant can invoke elenchus_start_session, elenchus_get_context, and subsequent verification commands. Then issue natural-language prompts like “Please verify src/auth for security issues” and let the system drive iterative analysis through the defined session lifecycle.

Prerequisites:

• Node.js v18 or later installed on your system
• Internet access to fetch packages

Installation steps:

1. Ensure Node.js is installed. You can verify with: node -v npm -v

2. You don’t need to manually install in advance if you’re using npx. The MCP client can run the server directly via npx. If you prefer a local installation, install the package: npm install -g @jhlee0409/elenchus-mcp

3. Verify installation (optional): npx -y @jhlee0409/elenchus-mcp --help

4. Configure your MCP client to use the server:

   • For Claude Desktop/Codes, add the mcpServers entry as shown in the Quick Start section of the README (npx -y @jhlee0409/elenchus-mcp).
   • Ensure your client can access your environment where Node.js is installed.

5. Run a verification session via your MCP client using elenchus_start_session (inputs like target, requirements, and workingDir according to your project).

Note: If you use a local environment, you may also pin a specific version of the MCP package to avoid breaking changes by using the package’s exact version in the npx command if needed.

Tips and considerations:

• The server depends on a multi-round Verifier↔Critic loop; you can tune evaluation via optional verification modes (standard/fast-track/single-pass) and differential/configuration hooks if supported by your client.
• If you enable LLM-based evaluation (llmEvalConfig), you can adjust convergence, severity, edge-case, and false-positive checks to suit your risk tolerance.
• Session management features (checkpointing, global storage, audit trails) help maintain traceability; ensure your client’s environment preserves these capabilities across sessions.
• Dependency analysis can cover multiple languages via tree-sitter; ensure your project’s codebase is accessible to the MCP client’s working directory.
• Common issues often relate to environment PATH problems for node, or mismatched working directories when targeting files; ensure workingDir accurately points to the project root for relative paths.