Owlex

What agentic-mcp-tools/owlex MCP server does

The agentic-mcp-tools/owlex MCP server gives an MCP-connected coding workflow access to several command-line AI agents. It can send a prompt to Codex, Gemini, OpenCode, ClaudeOR, and AiChat independently, collect their responses, and return a combined result. Claude can then synthesize the discussion.

The main council_ask operation supports an optional second round. In the first round, agents answer without seeing the other responses. In the next round, they can revise their positions after reviewing the collected answers. A critique mode changes the second-round behavior so agents challenge one another rather than simply revise.

How it works

A council request can include Claude's initial opinion, a timeout, specialist role assignments, or a predefined team. Supported roles include security, performance, skepticism, architecture, maintenance, developer experience, and testing. Team presets include security_audit, code_review, architecture_review, devil_advocate, balanced, and optimal.

The project also provides separate start and resume operations for each supported agent. Resuming uses the session identifier or continuation convention supported by that agent. Council work can run asynchronously, allowing the caller to continue other tasks and later use wait_for_task, get_task_result, list_tasks, or cancel_task.

Setup and configuration

Install the project with uv from its Git repository, then register the owlex-server executable in an MCP configuration file. The README shows a .mcp.json entry using that command. The documented Python requirement is version 3.10 or newer.

Configuration variables can exclude selected agents, choose a default team, share Claude's initial opinion automatically, and set the default timeout. Agent-specific settings control Codex approval bypassing, Gemini automatic approval, and whether OpenCode operates in plan or build mode. OPENROUTER_API_KEY enables the ClaudeOR agent, while CLAUDEOR_MODEL and AICHAT_MODEL select models for the relevant integrations.

Tools and capabilities

The agentic-mcp-tools/owlex MCP server documents these capability groups:

• Council queries with deliberation, critique, roles, teams, and timeouts.
• New and resumable sessions for Codex, Gemini, OpenCode, ClaudeOR, and AiChat.
• Background council task creation, inspection, waiting, and cancellation.
• Claude Code slash commands: /codex, /gemini, /council, and /critique.

The listed agent profiles have different intended strengths. Codex is described for reasoning and code review, Gemini for large-context and multimodal work, OpenCode for configurable alternative perspectives, ClaudeOR for models accessed through OpenRouter, and AiChat for multiple provider backends.

Limitations and notes

The council is not intended for every prompt; its purpose is to add comparison for decisions that warrant additional review. Codex and Gemini use existing subscriptions, while OpenCode and AiChat consume provider-dependent API tokens. Running multiple agents can therefore increase usage costs. Agents can be excluded with COUNCIL_EXCLUDE_AGENTS when cost or availability is a concern.

The CODEX_BYPASS_APPROVALS and GEMINI_YOLO_MODE settings can approve actions automatically, so they should be enabled only when that behavior is acceptable. OpenCode defaults to the read-oriented plan mode and can be changed to build.

The repository is released under the MIT license and identifies Python 3.10+ and MCP compatibility. The README does not state that every listed agent is available without separately installing or authenticating its underlying CLI or provider.