IntermediateEnglish16 lectures · 42m

Claude Code in Production: MCP, Headless, Agent SDK

Connect tools with MCP, build your own server, run Claude Code headless in scripts and CI, and program it with the Agent SDK

Watch 3 lectures free

What you'll learn

  • ✓Explain MCP tools, resources and prompts and add servers with the right scope
  • ✓Apply permission rules to MCP tools, and know why name-based deny lists leak
  • ✓Build an MCP server in Python with tools, resources and prompts, over stdio and HTTP
  • ✓Run Claude Code non-interactively with a fixed tool list and schema-constrained JSON output
  • ✓Turn an agent review into a CI gate that passes or fails on an exit code
  • ✓Program Claude Code from Python with the Agent SDK: custom tools, MCP and subagents

About this course

Interactive sessions are where you learn Claude Code. Production is where it connects to your tools, runs without a human at the keyboard, and becomes a component in your own software. This course covers those three moves.

Section 1 is MCP, the Model Context Protocol: what tools, resources and prompts are, how to add a server in one command, how scopes and the shared .mcp.json work, how project servers are approved, and how permission rules apply to MCP tools, including why a name-based deny list leaks. Section 2 builds your own MCP server in Python with the MCP SDK: tools whose contract comes from type hints and docstrings, resources and prompts (a server prompt appears as a slash command), and the same server moved onto HTTP so a team can share it. Section 3 is headless mode: claude -p, piping a diff in, restricting tools for unattended runs, schema-constrained JSON output, a quality gate that fails with an exit code, and a GitHub Actions workflow. Section 4 is the Agent SDK for Python: the query loop and options, in-process custom tools, MCP servers and subagents from code. The capstone builds a release-notes agent twice, first as a headless script around your own MCP server, then as an SDK program with a fact-checking subagent.

Every terminal lecture is a real recorded session, the Agent SDK lectures included. Resources contains every server and script from the course with pinned package versions.

**Prerequisite.** *Claude Code Hands-On* or equivalent daily use of Claude Code.

Curriculum7 sections · 16 lectures · 42m

Section 0. Welcome

Free preview

Section 1. MCP

Free preview

Section 2. Your Own MCP Server

  • 🔒Build your own MCP server with the Python SDK3:36
  • 🔒Resources and prompts, the two capabilities people skip2:30
  • 🔒HTTP transport - one server, many clients2:36

Section 3. Headless Mode

  • 🔒Headless mode - Claude Code as a command2:36
  • 🔒Permissions and structured output for unattended runs2:48
  • 🔒Putting the agent in continuous integration3:00

Section 4. The Agent SDK

  • 🔒The Agent SDK - Claude Code as a library2:24
  • 🔒Custom tools, MCP and subagents from code2:18

Section 5. Capstone

  • 🔒Capstone - a release-notes agent2:48
  • 🔒Capstone - the same agent as a program2:30

Section 6. Recap

  • 🔒Recap - what you built, and what to do next2:00

Requirements

  • · Claude Code installed and working (see Claude Code Hands-On)
  • · Python 3.10+, Node.js 18+ (for npx) and jq
  • · A Claude subscription for following along; an Anthropic API key for anything you deploy for other people

Who this is for

  • · Developers who use Claude Code daily and want it to reach their databases, trackers and internal tools
  • · Platform and DevOps engineers wiring an agent into scripts, CI and scheduled jobs
  • · Teams evaluating the Agent SDK for their own products

Read alongside the course

    The voice-over in this course is synthesized with a text-to-speech model from scripts written and reviewed by the instructor, and the on-screen material (notebooks, code, slides) is the instructor's own work.

    Claude Code in Production: MCP, Headless, Agent SDK | SOTAAZ Blog