MCP Server Integration

Every connector you’ve used in Cowork — Google Drive, Slack, GitHub — runs on the same technology: MCP. Understanding MCP turns you from a connector user into someone who can connect Cowork to anything.

🔄 Quick Recall: In the previous lesson, you built advanced skills with context files, conditional logic, and multi-step workflows. MCP adds the external data layer — your skills can now pull information from any connected tool.

What Is MCP?

MCP stands for Model Context Protocol. It’s an open standard — not owned by Anthropic, not specific to Claude — that provides a standardized way for AI tools to connect with external data sources.

Think of it like USB for AI. Before USB, every device had its own proprietary connector. MCP does the same thing for AI-to-tool connections: one standard protocol, any tool.

When you connect Google Drive to Cowork, what actually happens is:

1. An MCP server runs that speaks both “Google Drive API” and “MCP”
2. Cowork connects to that MCP server
3. Claude can now search, read, and write to Drive through the standard MCP interface

Every built-in connector is just a pre-configured MCP server. The 38+ connectors in Cowork? They’re 38 MCP servers that Anthropic or partners have already built.

Built-in vs Custom MCP

Most users never need custom MCP servers. The built-in connectors handle standard tools. But if your workflow depends on a tool that doesn’t have a built-in connector — your company’s internal API, a niche industry tool, a custom database — that’s when custom MCP comes in.

✅ Quick Check: What’s the difference between a built-in connector and a custom MCP server?

How MCP Works in a Plugin

When you build a plugin, MCP connections are defined in a .mcp.json file:

{
  "mcpServers": {
    "google-drive": {
      "url": "https://mcp.google.com/drive",
      "auth": "oauth2"
    },
    "internal-crm": {
      "url": "https://your-company.com/mcp-server",
      "auth": "api-key",
      "env": {
        "API_KEY": "CRM_API_KEY"
      }
    }
  }
}

This file sits at the root of your plugin directory. When the plugin loads, Cowork connects to each MCP server listed.

Your skills can then use these connections:

## Instructions
1. Connect to the CRM via the internal-crm connector
2. Pull the latest deals from the pipeline
3. Create a weekly sales report with deal status, values, and next steps

Claude knows how to use the MCP connection because the protocol is standardized — it doesn’t need to know the CRM’s specific API.

Anthropic’s Knowledge-Work Plugins and MCP

The 15 official plugins on GitHub include 40+ unique MCP connector configurations. Looking at the source code gives you real-world examples of how MCP is wired up:

• Marketing plugin: Google Analytics, HubSpot, Mailchimp connectors
• Engineering plugin: GitHub, Linear, Datadog connectors
• Sales plugin: Salesforce, Apollo, Outreach connectors
• Finance plugin: FactSet, Snowflake, QuickBooks connectors

Each plugin’s .mcp.json defines which connectors it needs. When you install the plugin, those connections are configured automatically — you just authorize them.

Using MCP With Your Skills

You don’t need to build custom MCP servers to benefit from MCP in your skills. You can reference existing connectors:

## Instructions

### Data Sources
This skill uses the following connectors:
- **Google Sheets** — for the team capacity spreadsheet
- **Jira** — for sprint data and ticket status
- **Slack** — for team standup messages

### Workflow
1. Pull current sprint data from Jira (tickets in progress, completed, blocked)
2. Read the team capacity spreadsheet from Google Sheets
3. Check #engineering-standup in Slack for today's updates
4. Generate the daily engineering dashboard

As long as the connectors are installed and authorized, the skill uses them automatically. You’re just telling Claude which connectors to use for which data.

When Custom MCP Makes Sense

You need a custom MCP server when:

• Your company has an internal tool with an API but no built-in connector
• You want to connect to a database directly (PostgreSQL, MongoDB, etc.)
• You need to integrate with a niche industry tool
• You want to automate interactions with a proprietary platform

Building a custom MCP server requires some technical ability — typically Python or Node.js, familiarity with APIs, and the ability to deploy a web server. This is a “work with your dev team” situation for most non-technical users.

Anthropic’s documentation at support.claude.com covers custom MCP server setup in detail. The short version: your server needs to implement the MCP protocol (there are open-source SDKs), expose endpoints that Claude can call, and handle authentication.

✅ Quick Check: When would a non-technical user need a custom MCP server? (When their workflow depends on an internal company tool or niche software that doesn’t have a built-in connector.)

MCP Security

MCP servers access your data. Security matters.

Built-in connectors: Authenticated via OAuth2 (Google, Microsoft, Slack). Your credentials are handled by Anthropic’s secure infrastructure. You control what data Claude can access through the standard permission prompts.

Custom MCP servers: You control the server, so you control the security. Best practices:

• Use API keys with minimal permissions (read-only when possible)
• Don’t expose MCP servers to the public internet without authentication
• Log all MCP requests for audit trails
• Rotate API keys regularly

Plugin MCP: When you install a community plugin with MCP connections, check what it connects to. A “Sales Report” plugin that wants access to your email is suspicious.

Key Takeaways

• MCP is the open protocol behind all Cowork connectors — like USB for AI tools
• Built-in connectors are pre-configured MCP servers — 38+ available
• Custom MCP servers extend Cowork to any tool with an API
• Plugin .mcp.json files define which connectors a plugin needs
• Your skills can reference MCP connectors to pull external data
• Security: verify what custom plugins connect to before installing

Up Next

You know how to build skills and connect them to external tools. But where do you find skills others have built? And how do you share yours? Next lesson: OpenClaw and the broader skill-sharing ecosystem.