Skip to content
CW101

cwork101 / glossary

Glossary

One sentence per term, in plain language. Where a term has a full entry, the definition links straight to it.

A

Adaptive thinkingthinking
Reasoning the model decides to do on its own, with no request parameter to enable it. Standard on every Claude 5 model. Full entry →
Agent SDKClaude Agent SDK
The TypeScript and Python SDK for building custom agents on Claude Code's engine, tools and permission model. Full entry →
Agent SkillsSkills · SKILL.md
A folder of instructions, scripts and resources an agent loads on demand when a task matches its description. Full entry →
Agent team
Several full Claude Code sessions orchestrated together, able to message one another. Full entry →
AI for Science program
An Anthropic grant scheme providing API credits to researchers. A funding programme, not a product. Full entry →
Artifact
A standalone, shareable output rendered beside the conversation — a document, chart, page or piece of code. Full entry →
Auto memorymemory
Learnings Claude Code saves across sessions on its own, without you writing them down. Full entry →

B

Batch APImessage batches
Asynchronous request processing at a 50% discount on input and output tokens, and the only route to 300K output tokens. Full entry →

C

Claude Consumption UnitCCU
A marketplace billing unit worth $0.01, used when Claude is invoiced through AWS Marketplace or Azure Marketplace. 100 CCU equals $1.00 of fees. Full entry →
Claude Science
A desktop research workbench that runs scientific analyses, searches 60+ databases and keeps provenance from data to publication. Full entry →
CLAUDE.md
A markdown file in a project root that Claude Code reads at the start of every session. Full entry →
Compactionsummarisation
Summarising part or all of a session's context so work can continue when the context window fills. Full entry →
Completion certificate
Recognition Anthropic Academy issues for finishing a course. Issued by Anthropic, not by this reference. Full entry →
Connectorintegration
A managed integration in the Claude apps, enabled through the interface and governable centrally. Full entry →
Context window
Everything the model can see at once — instructions, tool definitions, files read and the conversation so far. Full entry →

D

Deprecated
A model that still works but is no longer recommended, and has an assigned retirement date and a named replacement. Full entry →

E

Effort
A parameter controlling how much adaptive reasoning a model does. Defaults to high on Opus 5 and Sonnet 5 via the API and Claude Code. Full entry →
Extended thinking
The older, explicit thinking parameter (thinking.type: enabled). Supported on Haiku 4.5 and earlier; not on any Claude 5 model. Full entry →

F

Fast mode
A research-preview mode giving faster output on Opus 5 and Opus 4.8 at double the standard token price, on the first-party API only. Full entry →

H

Hook
A shell command Claude Code runs at a defined point in its execution. Deterministic, unlike an instruction. Full entry →

I

Inference geographydata residency
A request parameter pinning inference to a geography. inference_geo: "us" applies a 1.1x price multiplier. Full entry →

L

Legacy model
A model still available and supported, but no longer receiving updates and liable to be deprecated later. Full entry →
llms.txt
A plain-text index of a site, published for language models to discover its structure.

M

Managed Agents
Anthropic-hosted agents with a managed sandbox, billed on tokens plus $0.08 per session-hour of running time. Full entry →
Managed MCP
Organisational control over which MCP servers an organisation's users may connect to. Full entry →
MCP Apps
Interactive applications that run inside an AI client, defined by the Model Context Protocol beyond tools and resources. Full entry →
MCP client
Software that consumes MCP servers. Claude Code, the Claude desktop app and Agent SDK applications are all MCP clients. Full entry →
MCP server
A process exposing tools, resources and prompts over the Model Context Protocol. Full entry →
Model Context ProtocolMCP
An open client/server protocol letting AI applications connect to external tools and data through one interface. Full entry →
MTokper million tokens
One million tokens — the unit all Claude API pricing is quoted in.

P

Pinned snapshotmodel alias
A model ID fixed to one specific release. Every Claude model ID is one, including the dateless format used from the 4.6 generation onward. Full entry →
Plugin
An installable bundle distributed through a marketplace, able to carry skills, commands, hooks, MCP servers and subagents together. Full entry →
Progressive disclosure
How skills load in three stages — discovery, activation, execution — so many skills cost little context until one is needed. Full entry →
Prompt cachingcache_control
Reusing already-processed parts of a prompt across requests. A cache hit costs 0.1x the base input price. Full entry →
Prompt injection
Content in a page, file or message written to instruct the agent rather than inform the user. The central risk in computer use, browser use and CI. Full entry →

R

Rate limit429
A throughput cap on the API, set by usage tier. Throttles requests; does not cap spend. Full entry →
Reliable knowledge cutoff
The date through which a model's knowledge is most extensive and reliable — distinct from, and usually earlier than, the training data cutoff. Full entry →
ResearchDeep Research
A mode in the Claude apps where Claude runs many searches in parallel and returns a cited report rather than a single answer. Full entry →
Retired
A model that is no longer available. Requests to it fail. Full entry →
Routinescheduled task
A Claude Code task that runs in the cloud on a schedule, or on an API call or GitHub event. Full entry →

S

Server-side tool
A tool Anthropic hosts and executes, such as web search or code execution, sometimes carrying usage-based charges beyond tokens. Full entry →
Session
A unit of Claude Code work with its own context and history, movable between terminal, desktop, web and mobile. Full entry →
Subagentsub-agent
A child agent spawned by a session to handle a subtask in its own context window and return a result. Full entry →
Surface
One of the places a product runs — terminal, IDE, desktop, web, mobile, Slack, CI. Feature availability often differs between them. Full entry →
System cardmodel card
A document published with a model release reporting capability evaluations, safety testing, risks and known limitations. Full entry →

T

Tokenizer boundarytokenizer change
Claude Opus 4.7 and later produce roughly 30% more tokens for the same text than Sonnet 4.6 and earlier, which makes cross-generation price and context comparison misleading. Full entry →
Tool usefunction calling
The mechanism by which Claude calls functions you define, or tools Anthropic hosts, and continues from the results. Full entry →

U

Usage tier
The Start, Build or Scale band an API organisation sits in, determining its rate limits. Full entry →

W

Worktree
A separate git checkout giving each parallel Claude Code session its own files. Isolation stops at the filesystem — it is not a sandbox. Full entry →

Z

Zero data retentionZDR
A configuration in which submitted data is not retained after processing. Separate from data residency and from training defaults. Full entry →

Start typing to search every entity in the reference.