Ch 4: Tools — Letting the Model Act

Ch 4 — Tools: Letting the Model Act

Discovery, invocation, annotations, progress reporting, and the full tool lifecycle

Index Under the Hood →

High Level

Click play or press Space to begin...

Step- / 8

AWhat Are Tools?Functions the model can call to take action in the real world

psychology

LLM DecidesModel sees tool list
and chooses to call

calls

build

Toolsearch_issues,
run_query, send_msg

returns

output

ResultText, images, or
structured data

arrow_downward Step 1: Discover — what tools does the server have?

BTool DiscoveryClient asks the server what tools it offers

tools/listClient requests
available tools

returns

list

Tool ListName, description,
input schema

injected

psychology

LLM ContextTools become part
of system prompt

arrow_downward Step 2: Invoke — the model calls a tool

CTool InvocationLLM decides → host approves → server executes

psychology

LLM Callstool_call with
name + arguments

consent

shield

Host ApprovesCheck annotations
ask user if needed

tools/call

play_arrow

Server ExecutesRuns the function
returns result

arrow_downward Tool annotations: metadata that helps the host decide

DTool AnnotationsMetadata that tells the host how dangerous a tool is

menu_book

readOnlyHintDoesn't modify
anything (safe)

warning

destructiveHintDeletes or modifies
data (dangerous)

replay

idempotentHintSafe to retry
same result

arrow_downward Progress reporting: long-running tools

EProgress ReportingLong-running tools can report progress as they work

hourglass_top

Tool RunningLong operation
in progress

notify

trending_up

Progress25/100, 50/100...
via notifications

done

check_circle

Final ResultResponse with
content array

arrow_downward Dynamic tools: servers can add/remove tools at runtime

FDynamic Tool ListsServers can add or remove tools while running

add_circle

Tool AddedServer registers
new tool at runtime

notify

notifications

list_changedServer sends
change notification

refresh

Re-discoverClient calls
tools/list again