tutorpackSee prices
All courses

AI at work · course 28 of 105

MCP Agent Tools Built to Spec

Give an AI agent tools with the Model Context Protocol: valid tool definitions, schema checks, which error is which, reading results, and connecting servers in Claude Code.

4 modules

Try a question

Not on sale yet. When sales open: $4.99 for this course, paid once, or every course for $9.99 a month. Prices

Preview

What it teaches

How the Model Context Protocol (MCP) gives an AI agent tools, keyed to the current version, 2026-07-28: what makes a tool definition valid (name, inputSchema, tool-name rules, the new x-mcp-header), whether a call’s arguments pass the tool’s schema, which kind of error a failure should be (JSON-RPC protocol error vs a result with isError: true), how to read a 2026-07-28 tool result (resultType, structuredContent, outputSchema), and how to connect servers in Claude Code (claude mcp add, the -- separator, local / project / user scopes, .mcp.json entries, ${VAR} expansion). The AI asks 3 questions, places you, shows a short card per module, then gives you one definition, call, result or command at a time.

How long

46 items. About 25-80 rounds (about 25 if you know MCP, 50 to 80 for a beginner) plus a 12-item final: 2 sessions (estimate, not timed with a real student). The tutor gives a one-line save code at the end of each session.

The modules

4 modules, drilled one item at a time.

  1. M0

    What MCP is, version 2026-07-28

    What MCP is and what changed in 2026-07-28

  2. M1

    Read a tool definition

    Is this tool definition valid?

  3. M2

    Tool calls, errors and results

    Check a call’s arguments, name the error, read a result

  4. M3

    Connect servers in Claude Code

    Connect a server in Claude Code

Before it teaches, it asks

  1. 1
    What are you building with MCP (your own server, connecting servers to Claude Code, an agent product), and what have you used so far?
  2. 2
    Valid under MCP 2026-07-28? If not, what’s wrong?
    {
    “name”: “get_order”,
    “description”: “Look up an order by its ID”
    }
  3. 3
    Write the claude mcp add command for a remote HTTP server named notion at https://mcp.notion.com/mcp.

What’s exact, and what isn’t

Show

Straight from the course’s own notes.

Checked by script against the official schema
Every M1 and M2 key comes from a small validator that reads MCP’s own JSON schema for 2026-07-28 (a byte-identical copy of the file in the spec’s GitHub repo, downloaded 2026-10-10) plus the tool-name and x-mcp-header rules quoted from the spec’s Tools page. Error-code meanings come from the schema’s own definitions. Every item also has a hand-written answer the validator must agree with. The suggested fixes and the error kinds are written by hand; a test run read each of them against the spec.
Sourced
M0 and M3 keys quote the MCP spec (Tools, Transports, Versioning pages) and Claude Code’s MCP docs. 58 quotes matched word for word on 5 live pages (re-checked 2026-10-11). Some Claude Code setup answers are checked by script only for their quotes; a test run read them against the docs.
Honest limits
MCP changes fast. This course follows 2026-07-28 and Claude Code’s docs as of 2026-10-10; older guides (2025 and earlier) describe a handshake and sessions that 2026-07-28 replaced. One grey area is flagged, not tested: whether a call that fails its inputSchema is a protocol or an execution error (the spec’s two texts point different ways).
Not in it yet
Writing a server in an SDK, resources and prompts, authorization, the multi-round-trip input_required flow, other clients' config.