Claude Delegator

GPT expert subagents for Claude Code. Five specialists that can analyze AND implement—architecture, security, code review, and more.

Claude Delegator in action

Install

Inside a Claude Code instance, run the following commands:

Step 1: Add the marketplace

/plugin marketplace add jarrodwatts/claude-delegator

Step 2: Install the plugin

/plugin install claude-delegator

Step 3: Run setup

/claude-delegator:setup

Done! Claude now routes complex tasks to GPT experts automatically.

Note: Requires Codex CLI or Gemini CLI. Setup guides you through installation.

What is Claude Delegator?

Claude gains a team of GPT and Gemini specialists via native MCP. Each expert has a distinct specialty and can advise OR implement.

Note: You can use either provider (GPT or Gemini), or both. The plugin will automatically detect which one is configured and route tasks accordingly.

What You Get	Why It Matters
5 domain experts	Right specialist for each problem type
GPT or Gemini	Use your preferred model provider
Dual mode	Experts can analyze (read-only) or implement (write)
Auto-routing	Claude detects when to delegate based on your request
Synthesized responses	Claude interprets expert output, never raw passthrough

The Experts

Expert	What They Do	Example Triggers
Architect	System design, tradeoffs, complex debugging	"How should I structure this?" / "What are the tradeoffs?"
Plan Reviewer	Validate plans before you start	"Review this migration plan" / "Is this approach sound?"
Scope Analyst	Catch ambiguities early	"What am I missing?" / "Clarify the scope"
Code Reviewer	Find bugs, improve quality	"Review this PR" / "What's wrong with this?"
Security Analyst	Vulnerabilities, threat modeling	"Is this secure?" / "Harden this endpoint"

When Experts Help Most

Architecture decisions — "Should I use Redis or in-memory caching?"
Stuck debugging — After 2+ failed attempts, get a fresh perspective
Pre-implementation — Validate your plan before writing code
Security concerns — "Is this auth flow safe?"
Code quality — Get a second opinion on your implementation

When NOT to Use Experts

Simple file operations (Claude handles these directly)
First attempt at any fix (try yourself first)
Trivial questions (no need to delegate)

How It Works

You: "Is this authentication flow secure?"
                    ↓
Claude: [Detects security question → selects Security Analyst]
                    ↓
        ┌───────────────────────────────┐
        │  mcp__codex__codex            │
        │  (or mcp__gemini__gemini)     │
        │  → Security Analyst prompt    │
        │  → Expert analyzes your code  │
        └───────────────────────────────┘
                    ↓
Claude: "Based on the analysis, I found 3 issues..."
        [Synthesizes response, applies judgment]

Key details:

Each expert has a specialized system prompt (in prompts/)
Claude reads your request → picks the right expert → delegates via MCP (GPT or Gemini)
Responses are synthesized, not passed through raw
Experts can retry up to 3 times before escalating
Multi-turn conversations preserve context via threadId for chained tasks

Multi-Turn Conversations

For chained implementation steps, the expert preserves context across turns:

Turn 1: mcp__*__* → returns threadId
Turn 2: mcp__*__*-reply(threadId) → expert remembers turn 1
Turn 3: mcp__*__*-reply(threadId) → expert remembers turns 1-2

Use single-shot (codex or gemini only) for advisory tasks. Use multi-turn for implementation chains and retries.

Configuration

Operating Modes

Every expert supports two modes based on the task:

Mode	Sandbox	Use When
Advisory	`read-only`	Analysis, recommendations, reviews
Implementation	`workspace-write`	Making changes, fixing issues

Claude automatically selects the mode based on your request.

Configuration Defaults

Set global defaults in ~/.codex/config.toml instead of passing parameters on every call:

sandbox_mode = "workspace-write"
approval_policy = "on-failure"

Per-call parameters override these defaults. See Codex CLI docs for all config options.

Manual MCP Setup

If /setup doesn't work, register the MCP server(s) manually:

# For Codex (GPT)
# Idempotent: safe to rerun
claude mcp remove codex >/dev/null 2>&1 || true
claude mcp add --transport stdio --scope user codex -- codex -m gpt-5.3-codex mcp-server

# For Gemini
# Idempotent: safe to rerun
claude mcp remove gemini >/dev/null 2>&1 || true
claude mcp add --transport stdio --scope user gemini -- node ${CLAUDE_PLUGIN_ROOT}/server/gemini/index.js

Verify with:

claude mcp list
printf '{"jsonrpc":"2.0","id":"health","method":"initialize","params":{}}\n' | node ${CLAUDE_PLUGIN_ROOT}/server/gemini/index.js

Customizing Expert Prompts

Expert prompts live in prompts/. Each follows the same structure:

Role definition and context
Advisory vs Implementation modes
Response format guidelines
When to invoke / when NOT to invoke

Edit these to customize expert behavior for your workflow.

Requirements

You need at least one of the following providers configured:

Codex CLI (for GPT): npm install -g @openai/codex
Gemini CLI (for Gemini): npm install -g @google/gemini-cli

Authentication:

Codex: run codex login
Gemini: run gemini once and complete the sign-in flow (or set GOOGLE_API_KEY)

Commands

Command	Description
`/claude-delegator:setup`	Configure MCP server and install rules
`/claude-delegator:uninstall`	Remove MCP config and rules

Troubleshooting

Issue	Solution
MCP server not found	Restart Claude Code after setup
Provider not authenticated	Codex: run `codex login`. Gemini: run `gemini` once to complete sign-in (or set `GOOGLE_API_KEY`)
Tool not appearing	Run `claude mcp list` and verify registration
Expert not triggered	Try explicit: "Ask GPT to review..." or "Ask Gemini to review..."

Development

git clone https://github.com/jarrodwatts/claude-delegator
cd claude-delegator

# Test locally without reinstalling
claude --plugin-dir /path/to/claude-delegator

See CONTRIBUTING.md for guidelines.

Acknowledgments

Expert prompts adapted from oh-my-opencode by @code-yeongyu.

License

MIT — see LICENSE

claude-delegator