claude-agent-mcp

An MCP (Model Context Protocol) server that wraps the Claude Agent SDK. It uses Claude Code's built-in OAuth authentication, so no API key is required — just claude login and you're ready to go.

Features

| Tool | Description | |------|-------------| | claude_query | General-purpose queries — uses Claude's reasoning without any tools | | claude_code_task | Coding tasks with file system access (read/write/edit/search) | | claude_web_search | Web search with result summarization | | claude_multi_turn | Fully customizable agent execution — configure tools, permissions, budget, and more |

Prerequisites

• Bun v1.0+
• Claude Code CLI installed and logged in (claude login)

Installation

git clone https://github.com/j-token/claude-agent-mcp.git
cd claude-agent-mcp
bun install

Usage

Run directly

bun start

Add to Claude Code as an MCP server

Add the following to your Claude Code MCP settings (~/.claude/settings.json):

{
  "mcpServers": {
    "claude-agent": {
      "command": "bun",
      "args": ["run", "/path/to/claude-agent-mcp/src/index.ts"]
    }
  }
}

Add to Claude Desktop

Add to your Claude Desktop config (claude_desktop_config.json):

{
  "mcpServers": {
    "claude-agent": {
      "command": "bun",
      "args": ["run", "/path/to/claude-agent-mcp/src/index.ts"]
    }
  }
}

Tools

claude_query

Send a prompt to Claude without any tools. Best for reasoning, text generation, and Q&A.

| Parameter | Type | Required | Description | |-----------|------|----------|-------------| | prompt | string | Yes | The prompt to send | | systemPrompt | string | No | System prompt | | model | string | No | Model to use (e.g., claude-sonnet-4-6) | | maxTurns | number | No | Max conversation turns (default: 3) |

claude_code_task

Perform coding tasks with file system access. Can read, write, edit, and search files.

| Parameter | Type | Required | Description | |-----------|------|----------|-------------| | prompt | string | Yes | Description of the coding task | | cwd | string | No | Working directory (absolute path) | | permissionMode | string | No | plan (default), acceptEdits, or bypassPermissions | | allowedTools | string[] | No | Tools to allow (default: Read, Edit, Write, Glob, Grep, Bash) | | model | string | No | Model to use | | maxTurns | number | No | Max turns (default: 10) |

claude_web_search

Search the web and get a summarized result.

| Parameter | Type | Required | Description | |-----------|------|----------|-------------| | searchQuery | string | Yes | What to search for | | model | string | No | Model to use | | maxTurns | number | No | Max turns (default: 5) |

claude_multi_turn

Fully customizable agent execution. Configure every aspect of the agent run.

| Parameter | Type | Required | Description | |-----------|------|----------|-------------| | prompt | string | Yes | The prompt | | systemPrompt | string | No | System prompt | | allowedTools | string[] | No | Tools to allow | | permissionMode | string | No | default, acceptEdits, plan, or bypassPermissions | | cwd | string | No | Working directory | | model | string | No | Model to use | | maxTurns | number | No | Max conversation turns | | maxBudgetUsd | number | No | Max budget in USD |

How It Works

┌─────────────┐     stdio      ┌───────────────────┐     OAuth     ┌─────────────┐
│  MCP Client │ ◄────────────► │ claude-agent-mcp  │ ◄───────────► │  Claude API  │
│ (Claude Code │                │  (this server)    │               │             │
│  / Desktop)  │                └───────────────────┘               └─────────────┘
└─────────────┘

1. An MCP client (Claude Code, Claude Desktop, etc.) connects to this server via stdio.
2. The client calls one of the registered tools (claude_query, claude_code_task, etc.).
3. This server delegates the request to the Claude Agent SDK, which uses your existing OAuth session.
4. The response is streamed back and returned to the client.

Development

# Type-check
bun run typecheck

# Build
bun run build

License

MIT