@ask-llm/claude-mcp

@ask-llm/claude-mcp

MCP server for consulting Anthropic Claude Code CLI from Codex CLI, Cursor, OpenCode, and other MCP clients. It fills the reverse collaboration path: Claude can ask Codex through @ask-llm/codex-mcp, and Codex can ask Claude through @ask-llm/claude-mcp.

Install for Codex CLI

codex mcp add claude -- npx -y @ask-llm/claude-mcp

Then ask Codex to use ask-claude for an independent review or second opinion.

This provider is for Codex and other non-Claude hosts. Claude Code rejects nested Claude Code sessions; the unified orchestrator automatically suppresses the Claude provider when Claude Code is already the host.

Prerequisites

• Node.js 20 or newer.
• Claude Code CLI installed, on PATH, and authenticated. The implementation is validated against Claude Code 2.1.206 and requires a version that supports --safe-mode.

Tools

• ask-claude — ask Claude for analysis, with optional native sessionId, model override, and relative includeDirs.
• get-usage-stats — in-memory token and duration totals.
• ping — verify the MCP server and Claude CLI installation.

Safety boundary

Every consultation runs Claude Code with --safe-mode and an explicit Read,Glob,Grep tool list. Claude can inspect the current workspace and allowed relative directories, but cannot run shell commands or modify files. Prompts are sent through stdin rather than command-line arguments.

Models and configuration

• Default: opus (override with ASK_CLAUDE_MODEL).
• Fallback: sonnet (override with ASK_CLAUDE_FALLBACK_MODEL). Claude Code's native --fallback-model handles overload or availability failures.
• Timeout: 600000 ms (override with ASK_CLAUDE_TIMEOUT_MS, then GMCPT_TIMEOUT_MS).

The tool returns both human-readable text and a structured AskResponse with the actual model, native Claude session ID, and token usage when the CLI reports it.