Decide the Connections
An agent is only as safe as the connections you allow. Here you'll take a real-looking agent workflow, classify what each MCP server exposes, run the "Should I connect this server?" aid, and refuse the one connection that shouldn't be allowed.
Work through it
0 of 5 done
The scenario
Maple Valley College runs an advising agent (the host). Three MCP servers are available to connect:
- Catalog β official, runs locally; exposes course-catalog data to read.
- Ticketing β official; exposes a "create advising ticket" action.
- StudyTips β an unknown third-party, hosted remotely; exposes tools that can "read student records and post messages."
Your decisions
Fill in each row. Saved on this device until you reset.
| Server | Primitive (tool/resource/prompt) | Touches student data? | Allow? + why |
|---|---|---|---|
| Catalog (official, local) | |||
| Ticketing (official) | |||
| StudyTips (unknown, remote) |
Saved automatically on this device.
Reveal model answer
| Server | Primitive | Student data? | Decision |
|---|---|---|---|
| Catalog | Resource (read) | No | ALLOW β read-only, official, local; least privilege. |
| Ticketing | Tool (act) | Maybe (ticket content) | ALLOW with care β it can act, so require human sign-off before tickets send (Lesson 11). |
| StudyTips | Tools (read records + post) | Yes | DONβT ALLOW β unknown third party, remote, and tools that touch student records. Use a vetted alternative or de-identified data. |
Key takeaway: the deciding factors are who runs it, what it can do (tool vs. resource), local vs. remote, and whether it touches student data. StudyTips fails all four.
Takeaway: MCP connects; it doesnβt secure. The approval is your control β grant the least each task needs.