Ask Codex from Claude Code, or open a multi-model council

Who it is for

This guide is for people already on Claude Code who want the same question checked by Codex, Gemini, and other models. The official main path is installing from the repository with uv, then starting owlex-server in .mcp.json. It can also run structured deliberation: each model answers alone first, then revises after seeing the first round, and Claude synthesizes at the end.

It is for people who first open a Codex session, or who fire council_ask and inspect the result asynchronously. Tools the source names include start_codex_session, resume_codex_session, start_gemini_session, resume_gemini_session, plus wait_for_task, get_task_result, and list_tasks. Do not invent other model-start tool names that were not in the table.

First try in 3 minutes

The official install command is below. Prerequisites: uv can run on this machine, and the Codex / Gemini (or other) agents you will use already work locally. Config location is the project or user .mcp.json: mcpServers.owlex.command is owlex-server.

uv tool install git+https://github.com/agentic-mcp-tools/owlex.git

Repository page:

https://github.com/agentic-mcp-tools/owlex

The first try uses the official tool start_codex_session: create a new Codex session. Before calling, confirm you are starting a Codex that is already configured on this machine. Success is a session created, not a missing command. This starts an external model process and may go online; do not read it as offline.

Use cases

Case 1: ask Codex or Gemini alone. That is start_codex_session / start_gemini_session. Afterward you can use resume_codex_session (session ID or --last) or resume_gemini_session (index or latest). The source does not write message-field details. On success you can continue the same session.

Case 2: open a council. That is council_ask. The official sample prompt is Should I use a monorepo or multiple repos for 5 microservices? Options include claude_opinion, deliberate (default true), critique, roles, team, timeout (default 300 seconds). Council runs in the background: start it, then use list_tasks for status, or get_task_result for a non-blocking result; use wait_for_task when you must block. Before calling, confirm you want the same question sent to several models.

Limits

council sends the question to several agents. Round 2 also lets them see one another’s first-round answers. The source does not write cost, log retention, or how to delete a session. timeout defaults to 300 seconds. Do not invent OpenCode / ClaudeOR / AiChat start-tool names that were not in the table; the body only says you can query those models from Claude Code.