Build Your Own MCP Connector
Developers who want their assistant, whichever one, to reach internal tools, databases, and APIs
Prerequisites: Module 19 or equivalent. Comfortable in TypeScript or Python. An internal API or tool worth wrapping.
Last verified 2026-09-17
Learning objectives
- Explain what MCP standardises: tools, resources, and prompts behind one protocol
- Build an MCP server with the official SDK and define tools the model actually uses well
- Choose a transport (stdio for local, HTTP for remote) and handle authentication
- Test the connector end to end from your assistant: Claude Code and the Claude apps, Gemini CLI, ChatGPT developer mode, or a Copilot Studio agent
- Package and distribute the connector to a team
What's in this module
7 lessons · 7 exercises · ~8 min reading
Starts withWhy a protocol at all~1 minExecutive summary
You will learn how the Model Context Protocol works and how to wrap an internal tool or API as a connector any MCP-capable assistant can call. End state is a working MCP server you wrote yourself: tools defined, tested against the assistant you use, and installed for your team. After this module you can design tool definitions a model uses reliably, choose the right transport and auth for your environment, and ship internal connectors instead of waiting for vendors. This is also the MCP leg of Anthropic's Claude Certified Developer exam.
Sources
- Model Context Protocol documentation: https://modelcontextprotocol.io
- Anthropic, "Remote MCP servers", the hosted-server side, which is what a custom connector actually is: https://platform.claude.com/docs/en/agents-and-tools/remote-mcp-servers
- Anthropic, "MCP connector", connecting a server to Claude over the API: https://platform.claude.com/docs/en/agents-and-tools/mcp-connector
- Claude Help Center, "Build custom connectors via remote MCP servers", the same job from the product side, including the submission route: https://support.claude.com/en/articles/11503834-build-custom-connectors-via-remote-mcp-servers
- OpenAI Help Center, "Developer mode and MCP apps in ChatGPT", which plans get it, how an admin enables it, and how an app is created from a server: https://help.openai.com/en/articles/12584461-developer-mode-and-mcp-apps-in-chatgpt
- OpenAI, "Remote MCP servers" (Responses API tools), the
mcptool type, supported transports and approval settings: https://developers.openai.com/api/docs/guides/tools-connectors-mcp - Gemini CLI, "MCP servers with Gemini CLI", the
mcpServerssettings block, transports andgemini mcp add: https://geminicli.com/docs/tools/mcp-server/ - Google, "Function calling with the Gemini API", remote MCP servers through the Interactions API and the streamable-HTTP-only rule: https://ai.google.dev/gemini-api/docs/function-calling
- Microsoft Learn, "Connect your agent to an existing MCP server" (Copilot Studio), the onboarding wizard, transports and authentication options: https://learn.microsoft.com/en-us/microsoft-copilot-studio/mcp-add-existing-server-to-agent
- Microsoft Learn, "Build a plugin for a declarative agent from an MCP server" (Microsoft 365 Copilot), the Agents Toolkit route: https://learn.microsoft.com/en-us/microsoft-365/copilot/extensibility/build-mcp-plugins
- Anthropic Academy courses (Intro to MCP, MCP Advanced Topics): https://anthropic.skilljar.com
- Anthropic quickstarts: https://github.com/anthropics/claude-quickstarts
Prefer the live room?
This module also runs inside our in-person bootcamps and workshops in Copenhagen.
More in Track J
Building on the Claude API
The Claude API from first call to first agent: messages, system prompts, streaming, tool use, and structured output.
Claude Solution Architecture
How to design a Claude solution end to end: platform choice, architecture pattern, evaluation, cost model, and safety plan: the five decisions every deployment stands or falls on.