Onboard

Dashboard
Mission Path
The Crew
Practice
Workbench

Bridge

Leaderboard
Shop
Profile

Blog Terms and conditions Privacy policy

OnboardMission Path

All tracks

UNIT 03 · CHIEF ENGINEER · 27 STEPS · —

Hello, Forge tools and mcps.

MCP & Tools•Intermediate

Forge teaches you to give the model hands. Twenty-seven lessons across seven units: typed schemas, descriptions the agent actually picks, handlers with rich errors, idempotency, authorization, and MCPs an agent uses on its own.

YOUR RUN0 / 27

XP BANKED0 / 272

CAPSTONECapstone

DEBRIEF · WHAT YOU'LL CLEAR

By end of shift, you can

Echo

SIGNS OFF ON THIS UNIT

Read a model's mood from its first sentence
Pick the right role for the right turn

Course path

0 / 27 steps cleared

Continue!

The model needs hands

What is a tool, exactly?

Your first tool definition

When does the LLM call a tool?

NEXT · STEP 1

The model needs hands

Pick up where you left off. The next shift is queued and Echo's standing by.

+10XP~3min27left

Write examples that actually teach

Spec a JSON schema Echo will sign off on

Diagnose why a 'good' prompt drifts

Clear a capstone with hearts to spare

The trace: how an agent decides

More tools, more realistic parameters

The tool name is already half the description

Required vs optional: what blocks and what leaves room

Enums, min/max, format: the schema validates for you

What the tool RETURNS is also a contract

Descriptions the LLM actually reads. the heart of the Track

Bad description vs good description

Mark read vs side-effect: the agent acts differently

When NOT to make a tool (and let the LLM solve it)

Error handling: the handler that doesn't break the agent

Error messages that help the agent recover

Idempotency: calling twice isn't the same as calling twice

Logging for tools: the trace that saves you in production

Chaining tools: when one isn't enough

How many tools is too many? The catalog tradeoff

Caching results: free when the model asks twice

Stateful MCPs: the cart that remembers

Authorization: the agent is NOT admin

The tool the agent ignores

Failure gallery: what you'll debug in production

Capstone: an inventory MCP the agent actually uses

MCP signed

Continue!

The model needs hands

What is a tool, exactly?

Your first tool definition

When does the LLM call a tool?

The trace: how an agent decides

More tools, more realistic parameters

The tool name is already half the description

Required vs optional: what blocks and what leaves room

Enums, min/max, format: the schema validates for you

What the tool RETURNS is also a contract

Descriptions the LLM actually reads. the heart of the Track

Bad description vs good description

Mark read vs side-effect: the agent acts differently

When NOT to make a tool (and let the LLM solve it)

Error handling: the handler that doesn't break the agent

Error messages that help the agent recover

Idempotency: calling twice isn't the same as calling twice

Logging for tools: the trace that saves you in production

Chaining tools: when one isn't enough

How many tools is too many? The catalog tradeoff

Caching results: free when the model asks twice

Stateful MCPs: the cart that remembers

Authorization: the agent is NOT admin

The tool the agent ignores

Failure gallery: what you'll debug in production

Capstone: an inventory MCP the agent actually uses

MCP signed

CAPSTONE IN 27 STEPS

Echo waits on the bridge

3 runs · 4/4 criteria. No misses.

0% completed

CHIEF ENGINEER

What you'll be able to do by the end of this track

Continue · Step 1 → The model needs hands

Start step