Level 6 of 14 · about 4 weeks

Tools, MCP and computer use

Connect agents to real systems through standard protocols, and keep those connections safe.

Modules

1. 1

   Tool design and least privilege

   Narrow tools, read-only by default, scoped credentials, and the controls that still work when the model is fooled

   Core 4 parts · about 5 hours

2. 2

   MCP architecture

   Hosts, clients and servers; tools, resources and prompts; and what changed when the protocol became stateless

   Core 5 parts · about 5 hours

3. 3

   MCP transports

   stdio and Streamable HTTP, progress and cancellation, notifications, broken connections, and what an exposed endpoint must check

   Core 6 parts · about 6 hours

4. 4

   Sampling and elicitation

   how a server asks the user or the host's model for something in the middle of a call, how that exchange is carried safely, and who actually controls it

   Common 5 parts · about 6 hours

5. 5

   MCP authorisation and security

   tokens, scopes and audiences, the authorization flow, the attacks on it, and the trust problems that no token can solve (poisoned tools and untrusted content)

   Core 6 parts · about 8 hours

6. 6

   Build and test a server

   a help-desk MCP server on a real database, and the test layers that tell you whether it is correct, safe and readable by a model

   Core 6 parts · about 9 hours

7. 7

   Agent-to-agent protocols

   A2A 1.0 from the card to the task, built as a small server and client, and what changes when the thing you call is another agent

   Common 5 parts · about 6 hours

8. 8

   Code execution and sandboxes

   letting an agent run the code it writes, with real operating-system limits, and the tests that show whether the walls hold

   Core 5 parts · about 8 hours

9. 9

   Browser and computer-use agents

   an agent that reads pages and clicks, three ways of looking at the same page, and the layers that limit what a hidden instruction can do

   Specialist 8 parts · about 8 hours