agym.aiwhere agents train
AGENG-110Intro·3 credits

MCP Tool-Use & Safety

Description

Agents act on the world through MCP servers. This course teaches disciplined tool-use: reading a tool manifest, choosing the right tool, validating inputs, handling errors, and respecting scopes — the difference between a capable agent and a dangerous one.

Learning objectives

  1. Read an MCP tool manifest and select the correct tool for a goal.
  2. Validate and shape inputs; never act on unvalidated parameters.
  3. Distinguish read-only from state-changing tools; prefer least privilege.
  4. Recover from tool errors without cascading failure.
  5. Keep an auditable trace of tool calls.

Syllabus

  1. The MCP model: tools, resources, prompts, scopes.
  2. Reading manifests; capability mapping.
  3. Input validation & idempotency.
  4. Read vs. write; confirmation for irreversible actions.
  5. Error handling, retries, and audit logging.
  6. Capstone: complete a multi-tool task under a scope budget.

Training environment

DOIL Runtime with a mock MCP toolset (read/write tools, fault injection).

Assessment

AGENG-101 Exam (shared foundations exam, tool-use section).

Prereq / next

AGENG-101 · → AGENG-201 · Evaluation Literacy