MCP Protocol

Full Model Context Protocol (MCP) support - both as server and client.

Overview

MCP is an open standard by Anthropic for connecting AI models to external tools and data sources. Quark Agent supports MCP in both directions:

• As Server: Expose quark-agent's tools to any MCP-compatible client (Claude Desktop, Cursor, etc.)
• As Client: Connect to external MCP servers and aggregate their tools

Server Mode

Expose your tools via JSON-RPC 2.0 over stdio or SSE:

```ts import { createQuarkMCPServer } from "quark-agent";

const mcp = await createQuarkMCPServer();

// stdio transport (for Claude Desktop integration) await mcp.runStdio();

// or SSE transport (for web-based clients) mcp.runSSE(3000); ```

Registered Tools

The server pre-registers these built-in tools:

Tool	Description
web_search	Search the web
web_fetch	Fetch a URL and return content
computer_screenshot	Take a screenshot

Claude Desktop Integration

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

json { "mcpServers": { "quark-agent": { "command": "npx", "args": ["github:Madapexai/quark-agent", "mcp-server"] } } }

Client Mode

Connect to external MCP servers and aggregate their tools:

```ts import { MCPRegistry } from "quark-agent";

const registry = new MCPRegistry();

// Connect via stdio (spawn subprocess) await registry.connect({ name: "filesystem", command: "npx", args: ["@modelcontextprotocol/server-filesystem", "/tmp"], });

// Connect via SSE await registry.connect({ name: "remote-tools", url: "https://example.com/mcp/sse", });

// List all aggregated tools const tools = registry.listAllTools(); // -> [{ server: "filesystem", tool: { name: "read_file", ... } }, ...]

// Call a tool const result = await registry.callTool("filesystem", "read_file", { path: "/tmp/test.txt" }); ```

API Endpoints

Method	Path	Description
GET	/api/mcp/tools	List built-in + external tools
GET	/api/mcp/servers	List connected servers
POST	/api/mcp/connect	Connect to a server
POST	/api/mcp/disconnect	Disconnect from a server

Connect via API

bash curl -X POST http://localhost:8788/api/mcp/connect \ -H "Content-Type: application/json" \ -H "Authorization: Bearer <token>" \ -d '{"name":"filesystem","command":"npx","args":["@modelcontextprotocol/server-filesystem","/tmp"]}'