Claude Octopus

One brain, many arms.

An MCP server that wraps the Claude Agent SDK, letting you run multiple specialized Claude Code agents — each with its own model, tools, system prompt, and personality — from any MCP client.

Why

Claude Code is powerful. But one instance does everything the same way. Sometimes you want a strict code reviewer that only reads files. A test writer that defaults to TDD. A cheap quick helper on Haiku. A deep thinker on Opus.

Claude Octopus lets you spin up as many of these as you need. Same binary, different configurations. Each one shows up as a separate tool in your MCP client.

Prerequisites

• Node.js >= 18

• Claude Code — the Claude Agent SDK is bundled as a dependency, but it spawns Claude Code under the hood, so you need a working claude CLI installation

• Anthropic API key (ANTHROPIC_API_KEY env var) or an active Claude Code OAuth session

Install

npm install claude-octopus

Or skip the install entirely — use npx directly in your .mcp.json (see Quick Start below).

Quick Start

Add to your .mcp.json:

{
  "mcpServers": {
    "claude": {
      "command": "npx",
      "args": ["claude-octopus@latest"],
      "env": {
        "CLAUDE_PERMISSION_MODE": "bypassPermissions"
      }
    }
  }
}

This gives you two tools: claude_code and claude_code_reply. That's it — you have Claude Code as a tool.

Multiple Agents

The real power is running several instances with different configurations:

{
  "mcpServers": {
    "code-reviewer": {
      "command": "npx",
      "args": ["claude-octopus@latest"],
      "env": {
        "CLAUDE_TOOL_NAME": "code_reviewer",
        "CLAUDE_SERVER_NAME": "code-reviewer",
        "CLAUDE_DESCRIPTION": "Strict code reviewer. Finds bugs and security issues. Read-only.",
        "CLAUDE_MODEL": "opus",
        "CLAUDE_ALLOWED_TOOLS": "Read,Grep,Glob",
        "CLAUDE_APPEND_PROMPT": "You are a strict code reviewer. Report real bugs, not style preferences.",
        "CLAUDE_EFFORT": "high"
      }
    },
    "test-writer": {
      "command": "npx",
      "args": ["claude-octopus@latest"],
      "env": {
        "CLAUDE_TOOL_NAME": "test_writer",
        "CLAUDE_SERVER_NAME": "test-writer",
        "CLAUDE_DESCRIPTION": "Writes thorough tests with edge case coverage.",
        "CLAUDE_MODEL": "sonnet",
        "CLAUDE_APPEND_PROMPT": "Write tests first. Cover edge cases. TDD."
      }
    },
    "quick-qa": {
      "command": "npx",
      "args": ["claude-octopus@latest"],
      "env": {
        "CLAUDE_TOOL_NAME": "quick_qa",
        "CLAUDE_SERVER_NAME": "quick-qa",
        "CLAUDE_DESCRIPTION": "Fast answers to quick coding questions.",
        "CLAUDE_MODEL": "haiku",
        "CLAUDE_MAX_BUDGET_USD": "0.02",
        "CLAUDE_EFFORT": "low"
      }
    }
  }
}

Your MCP client now sees three distinct tools — code_reviewer, test_writer, quick_qa — each purpose-built.

Agent Factory

Don't want to write configs by hand? Add a factory instance:

{
  "mcpServers": {
    "agent-factory": {
      "command": "npx",
      "args": ["claude-octopus@latest"],
      "env": {
        "CLAUDE_FACTORY_ONLY": "true",
        "CLAUDE_SERVER_NAME": "agent-factory"
      }
    }
  }
}

This exposes a single create_claude_code_mcp tool — an interactive wizard. Tell it what you want ("a strict code reviewer that only reads files") and it generates the .mcp.json entry for you, listing all available options you can customize.

In factory-only mode, no query tools are registered — just the wizard. This keeps routing clean: the factory creates agents, the agents do work.

Tools

Each non-factory instance exposes:

Per-invocation parameters (override server defaults):

Configuration

All configuration is via environment variables in .mcp.json. Every env var is optional.

Identity

Agent

Prompts

Advanced

Authentication

Leave both unset to inherit auth from the parent process. Set one per agent to use a different account or billing source.

Lists accept JSON arrays when values contain commas: ["path,with,comma", "/normal"]

Security

• **Permission mode defaults to **default — tool executions prompt for approval unless you explicitly set bypassPermissions.

• cwd overrides preserve agent knowledge — when the host overrides cwd, the agent's configured base directory is automatically added to additionalDirectories so it retains access to its own context.

• Tool restrictions narrow, never widen — per-invocation tools intersects with the server restriction (can only remove tools, not add). disallowedTools unions (can only block more).

• _reply** tool respects persistence** — not registered when CLAUDE_PERSIST_SESSION=false.

Architecture

┌─────────────────────────────────┐
│  MCP Client                     │
│  (Claude Desktop, Cursor, etc.) │
│                                 │
│  Sees: code_reviewer,           │
│        test_writer, quick_qa    │
└──────────┬──────────────────────┘
           │ JSON-RPC / stdio
┌──────────▼──────────────────────┐
│  Claude Octopus (per instance)  │
│                                 │
│  Env: CLAUDE_MODEL=opus         │
│       CLAUDE_ALLOWED_TOOLS=...  │
│       CLAUDE_APPEND_PROMPT=...  │
│                                 │
│  Calls: Agent SDK query()       │
└──────────┬──────────────────────┘
           │ in-process
┌──────────▼──────────────────────┐
│  Claude Agent SDK               │
│  Runs autonomously: reads files,│
│  writes code, runs commands     │
│  Returns result + session_id    │
└─────────────────────────────────┘

How It Compares

Development

pnpm install
pnpm build       # compile TypeScript
pnpm test        # run tests (vitest)
pnpm test:coverage  # coverage report

License

ISC - Xiaolai Li