DEV Community

curatedmcp for CuratedMCP

Posted on • Originally published at curatedmcp.com

Anthropic Claude MCP: Build Multi-Agent Workflows Inside Claude

Install guide and config at curatedmcp.com

Anthropic Claude MCP: Build Multi-Agent Workflows Inside Claude

What It Does

The Anthropic Claude MCP server lets you invoke Claude models as tools from within another Claude session—turning a single agent into a coordinated multi-agent system. Instead of routing between separate API calls or third-party services, you can now nest Claude instances with specialized roles, each running its own system prompt, temperature settings, and model choice.

This unlocks clean architectural patterns: an orchestrator Claude that breaks down complex tasks, delegates to a Haiku instance for quick triage, spins up Opus for deep reasoning on specific subtasks, and chains results back together—all without leaving your conversation.

The server supports Haiku, Sonnet, and Opus, giving you granular control over speed-vs-capability tradeoffs. Built-in prompt caching keeps token costs reasonable when reusing the same system prompts across multiple calls. You get full parameter control: temperature, max_tokens, and streaming support for long-form outputs.

How to Install

npx -y @anthropic-ai/mcp
Enter fullscreen mode Exit fullscreen mode

Add to your Claude Desktop config:

{
  "mcpServers": {
    "anthropic-claude-mcp": {
      "command": "npx -y @anthropic-ai/mcp"
    }
  }
}
Enter fullscreen mode Exit fullscreen mode

Once installed, Claude gains access to the Claude call tool and can invoke sub-agents directly within the conversation.

Real-World Use Cases

  • Code review + refactoring pipeline: Main Claude generates initial code, then spins up a second Claude with a "code critic" system prompt to review for bugs, performance, and style—both running in parallel, results merged back for refinement.

  • Legal document analysis: Route contract review to a Claude instance with legal domain expertise (via system prompt), while the orchestrator handles document triage, summarization, and risk flagging—specialised personas for specialized tasks.

  • Parallel research and synthesis: Split a research task across multiple Claude calls (market analysis, competitive landscape, technical deep-dive) running simultaneously, then have the main agent synthesize findings into a cohesive report.


Full install guides for Claude Desktop, Cursor, Windsurf, and more at CuratedMCP.

Top comments (0)