claude-delegator
Delegate tasks to Codex GPT 5.2 directly from within Claude Code.
claude mcp add --transport stdio jarrodwatts-claude-delegator node path/to/claude-delegator/server/gemini/index.js \ --env GOOGLE_API_KEY="your Gemini API key (if using Gemini)" \ --env CLAUDE_PLUGIN_ROOT="path to Claude Delegator plugin root"
How to use
Claude Delegator exposes a managed MCP server within Claude Code that routes complex Claude Code tasks to a team of expert GPT and/or Gemini subagents. The plugin provides five specialist roles—Architect, Plan Reviewer, Scope Analyst, Code Reviewer, and Security Analyst—that can either analyze a request or implement changes directly. You can choose to leverage GPT (Codex) or Gemini, or both, and Claude will auto-route tasks to the appropriate expert based on the prompt and your configuration. Use multi-turn conversations to chain implementation steps, where the expert passes a threadId back to continue context across turns. This MCP server is designed to help with architecture decisions, security reviews, code quality improvements, and thorough reviews before implementing changes. To enable it, install the plugin in Claude Code, then configure MCP if needed and run the setup command to start delegating complex tasks to the experts.
How to install
Prerequisites:
- Access to Claude Code with marketplace plugin installation capabilities
- Either Codex CLI (for GPT) or Gemini CLI (for Gemini) installed globally on your system
- Node.js and npm (if using the Node-based MCP server endpoint)
Install steps:
-
Add the Claude Delegator marketplace plugin inside Claude Code:
- Run: /plugin marketplace add jarrodwatts/claude-delegator
-
Install the plugin:
- Run: /plugin install claude-delegator
-
Run the MCP setup for the server:
- Run: /claude-delegator:setup
-
If you need manual MCP setup (alternative):
- For Codex (GPT): claude mcp add --transport stdio --scope user codex -- codex -m gpt-5.3-codex mcp-server
- For Gemini: claude mcp add --transport stdio --scope user gemini -- node /path/to/claude-delegator/server/gemini/index.js
-
Verify registration:
- Run: claude mcp list
- You can also send a health payload to the Gemini endpoint as shown in the README example.
Notes:
- Ensure your Gemini API key is available (GOOGLE_API_KEY) if using Gemini.
- If you run into authentication or server-not-found issues, restart Claude Code after setup and verify MCP registration with claude mcp list.
Additional notes
Tips and common issues:\n- Environment variables: set CLAUDE_PLUGIN_ROOT to the plugin directory; set GOOGLE_API_KEY if using Gemini.\n- If you switch providers (GPT or Gemini), Claude will auto-detect the configured provider and route tasks accordingly.\n- For multi-turn implementation flows, use the threadId mechanism described in the docs to preserve context across turns.\n- If the expert is not triggered, try explicit prompts such as 'Ask GPT to review...' or 'Ask Gemini to review...'.\n- Ensure the MCP server is registered under the correct transport (stdio) and scope (user or org) as per your workflow.\n- The Prompts for each expert live under prompts/ and can be customized if you fork the repo.
Related MCP Servers
awesome-claude-skills
A curated list of awesome Claude Skills, resources, and tools for customizing Claude AI workflows
deepcontext
DeepContext is an MCP server that adds symbol-aware semantic search to Claude Code, Codex CLI, and other agents for faster, smarter context on large codebases.
sandboxed.sh
Self-hosted orchestrator for AI autonomous agents. Run Claude Code & Open Code in isolated linux workspaces. Manage your skills, configs and encrypted secrets with a git repo.
sub-agents
Define task-specific AI sub-agents in Markdown for any MCP-compatible tool.
ask-user-questions
Better 'AskUserQuestion' - A lightweight MCP server/OpenCode plugin/Agent Skills + CLI tool that allows your LLMs ask questions to you. Be the human in the human-in-the-loop!
local-skills
Universal MCP server enabling any LLM or AI agent to utilize expert skills from your local filesystem. Reduces context consumption through lazy loading. Works with Claude, Cline, and any MCP-compatible client.