jarrodwatts-claude-delegator

Claude Delegator

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

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: