MCP

Interoperate

The Model Context Protocol (MCP) is a standard way for agents and tools to talk across applications. Mastra speaks it both directions: as a client it consumes tools exposed by external MCP servers; as a server it exposes your tools and agents to other MCP-aware apps (IDEs, other agents). This is how a Mastra app plugs into the ecosystem instead of living on an island.

graph LR
  EXT["External MCP servers<br/>(GitHub, DBs, …)"] -->|"client: consume tools"| M["Mastra app"]
  M -->|"server: expose tools / agents"| APPS["Other MCP-aware apps<br/>(IDEs, agents)"]

As a client — consume external tools

Configure an MCPClient with the servers you want, fetch their tools, and hand them to an agent like any local tool. The external server’s capabilities become available to the model.

import { MCPClient } from '@mastra/mcp';
import { Agent } from '@mastra/core/agent';

const mcp = new MCPClient({
  servers: {
    github: { url: new URL('https://example.com/mcp/github') },
  },
});

export const devAgent = new Agent({
  id: 'dev-agent',
  name: 'Dev Agent',
  instructions: 'Use the available tools to inspect repositories.',
  model: 'openai/gpt-4o',
  tools: await mcp.getTools(),
});

As a server — expose your tools

Wrap your tools (and agents) in an MCPServer so other MCP-aware clients can discover and call them. Your Mastra app becomes a building block in someone else’s agent.

import { MCPServer } from '@mastra/mcp';
import { weatherTool } from './tools/weather-tool';

export const server = new MCPServer({
  name: 'weather-server',
  version: '1.0.0',
  tools: { weatherTool },
});

Note

The @mastra/mcp API (MCPClient / MCPServer options, transport config) is version-sensitive — the dossier at _Sources/ai-system/mastra/ pins the exact version. Treat these as the shape of the API and confirm against what you install.

---

Reference: MCP overview · MCPClient · MCPServer

Next: Deployment — take the whole thing to production.