Claude AI Glossary: 100 Terms Every Enterprise User Must Know

An AI system that takes actions in response to goals rather than simply generating text. In the Claude context, an agent reads inputs, reasons about them, calls tools, and produces outputs — often repeating this loop multiple times before completing a task. Claude Code, Claude Cowork, and custom Claude API deployments can all operate as agents.

The iterative cycle in which an AI agent perceives its environment, reasons about what to do next, takes an action (often via a tool call), observes the result, and repeats. Enterprise AI agents run agentic loops that may span dozens of steps before completing a workflow.

The AI safety company that develops and maintains Claude. Founded in 2021. Valued at $380 billion as of 2026. Anthropic's research focus on Constitutional AI and interpretability shapes how Claude behaves across all products and plans.

Anthropic's official learning platform, hosted on Skilljar. Offers 13 free courses covering the Claude API, prompt engineering, Claude Code, MCP, and agent development. Required preparation for the Claude Certified Architect (CCA) exam.

A secret credential used to authenticate requests to the Claude API. Enterprise teams should store API keys in secrets management systems (AWS Secrets Manager, HashiCorp Vault) — never in code repositories or environment files committed to source control.

In Claude.ai and Cowork, an Artifact is a generated piece of content — code, a document, a data table, a rendered component — that is presented in a dedicated panel separate from the conversation. Artifacts can be iterated on, downloaded, or embedded.

A tamper-evident record of every action Claude takes within an enterprise deployment — what was asked, what tools were called, what files were accessed, and what outputs were produced. Required for compliance in regulated industries. Available in Claude Enterprise and via the API with custom logging.

Anthropic's asynchronous processing endpoint that handles large volumes of Claude API requests at 50% lower cost than the synchronous API. Ideal for overnight processing jobs, large document batches, and high-volume evaluation runs. Requests are submitted as a batch and results are retrieved when processing completes.

An HTTP header (anthropic-beta) used to opt into features that are available but not yet stable in the Claude API. Extended thinking, computer use, and certain prompt caching configurations have historically required beta headers before reaching GA.

Anthropic's official architecture certification. A proctored 60-question, 120-minute exam covering 5 domains: Claude API & application architecture, Model Context Protocol, Claude Code, agentic architecture, and safety & governance. Launched March 12, 2026. The first Claude-specific credential with the rigour of a cloud solutions architect exam.

Anthropic's family of AI models. Available in three tiers: Opus (highest capability, highest cost), Sonnet (balanced performance and cost, the workhorse for most enterprise applications), and Haiku (fastest and cheapest, suitable for high-volume lightweight tasks). Current generation: Opus 4.6, Sonnet 4.6, Haiku 4.5.

Anthropic's agentic coding tool for developers. Runs in the terminal. Has access to the file system, can execute commands, read and write code, call MCP servers, and delegate to sub-agents. Configured via CLAUDE.md files. Supports Skills (reusable capability bundles) and Hooks (event-driven governance triggers).

Anthropic's desktop AI agent for knowledge workers. Runs as a native application. Can read local files, connect to external services via Connectors, execute multi-step workflows, and be controlled remotely via Claude Dispatch. Deployed in Pro ($20/mo), Max ($100/mo), and Enterprise tiers.

Anthropic's mobile control interface for Claude Cowork. Allows users to trigger Cowork sessions, send instructions, and receive results via a smartphone app. Pairs with the desktop Cowork installation via QR code. Enables async AI workflows initiated on mobile.

Anthropic's organisation-level Claude plan. Includes SSO, admin controls, expanded context windows, audit logging, usage dashboards, and negotiated pricing for large teams. Designed for organisations deploying Claude across hundreds or thousands of users with compliance requirements.

Anthropic's certified partner programme for consulting firms, SIs, and ISVs building on Claude. Launched 2026 with $100M invested. Partners receive early access to features, joint go-to-market support, training resources, and listing in Anthropic's partner directory. We are a Claude Partner Network member.

A configuration file placed at the root of a project (or in a home directory) that provides Claude Code with persistent instructions, project context, allowed commands, architectural guidance, and tool configurations. The most important single file in a Claude Code enterprise deployment.

A Claude API capability that allows Claude to control a desktop environment — clicking buttons, typing text, reading screen contents, navigating applications — by issuing computer interaction commands. Available via the API as a beta feature. Used to automate workflows in legacy desktop applications that lack modern APIs.

In Claude Cowork, a pre-built integration that connects Claude to an external service — Gmail, Google Drive, Slack, DocuSign, FactSet, and others. Connectors handle authentication, data access, and permission scoping. Enterprise teams can request custom connectors through Anthropic's partner programme.

Anthropic's approach to training Claude to be helpful, harmless, and honest. Rather than relying solely on human feedback, Constitutional AI uses a set of principles to guide model behaviour, including self-critique and revision during training. It's what makes Claude's safety properties more consistent than pure RLHF approaches.

The total number of tokens Claude can process in a single API call, including the system prompt, conversation history, tool definitions, and the model's own output. Claude's context window is measured in tokens (approximately 0.75 words per token). Larger context windows allow processing longer documents and maintaining longer conversations.

A Claude capability (available in Claude.ai and Cowork) that conducts extended, multi-step web research with citations. Deep Research browses multiple sources, synthesises findings, and produces referenced reports. Enterprise use cases include market analysis, regulatory monitoring, and competitive intelligence.

The process by which an orchestrating Claude agent assigns a subtask to a sub-agent with a specific, scoped set of instructions and tools. Delegation enables parallel execution, context isolation, and specialisation within multi-agent systems.

Vector representations of text that capture semantic meaning. Anthropic does not currently offer a standalone embeddings API (use OpenAI or Cohere for embeddings); however, Claude's text understanding is informed by the same principles. Embeddings from third-party providers are used in RAG systems paired with Claude.

A Claude API mode that allows the model to perform extended internal reasoning before producing a response. Extended thinking improves performance on complex analytical tasks — multi-step maths, architecture decisions, legal analysis — at the cost of increased latency and token usage. Enabled via a beta header.

A prompting technique in which 2–10 example input-output pairs are included in the prompt to demonstrate the desired behaviour or format. More reliable than zero-shot for structured extraction tasks. Include examples that cover edge cases, not just the easy cases.

See Tool Use. In the Claude API, the mechanism is formally called "tool use," but the underlying pattern — defining structured functions Claude can invoke — is equivalent to function calling in other AI APIs.

Anthropic's fastest and most cost-efficient Claude model. Current version: Haiku 4.5. Suitable for high-volume tasks where speed and cost matter more than maximum capability: classification, simple extraction, customer service triage, and real-time interactive applications.

In Claude Code, an event-driven trigger that executes a specified command at a defined point in Claude's operation: before writing a file, before running a shell command, after generating code, before committing. Hooks are the primary mechanism for enforcing governance policies in enterprise Claude Code deployments.

A unit of text sent to the Claude API — roughly equivalent to 0.75 words. The system prompt, conversation history, tool definitions, and user messages all consume input tokens. Input tokens are priced separately from output tokens and are generally cheaper.

Research into understanding what happens inside AI models — what features they detect, how they reason, and why they produce specific outputs. Anthropic's interpretability research is foundational to building trust in Claude for high-stakes enterprise applications.

An open protocol developed by Anthropic that standardises how AI models connect to external tools, data sources, and services. MCP replaces ad-hoc API integrations with a consistent server-client architecture. Enterprise teams build MCP servers to connect Claude to internal databases, CRMs, ticketing systems, and proprietary APIs. See our MCP Protocol Guide.

A service that exposes tools and resources to Claude via the Model Context Protocol. An MCP server defines what Claude can access and do: query a database, search a document store, post to a Slack channel, retrieve a JIRA ticket. Enterprise MCP servers are built with authentication, rate limiting, and audit logging appropriate to internal systems.

Persistence of information between conversations or tasks. Claude does not retain memory across conversations by default; memory must be explicitly managed — either by maintaining conversation history in the context, using external storage via MCP servers, or using Claude.ai's built-in memory features in Pro/Max plans.

Anthropic's primary API endpoint for interacting with Claude. Accepts a list of message objects (alternating human/assistant turns), a system prompt, model selection, and parameters. Returns a completion response. The foundation of all Claude API integrations.

An architecture in which multiple Claude agent instances collaborate on a shared task. A typical pattern: one orchestrator agent decomposes the goal, several specialist sub-agents execute components in parallel, and the orchestrator synthesises results. Enables complex enterprise workflows that exceed single-agent context or capability constraints.

Anthropic's commitment that API inputs and outputs are not used to train future Claude models by default (unlike Claude.ai consumer free tier). Enterprise customers using the API or Claude Enterprise have a clear contractual non-training guarantee — a critical requirement for processing proprietary business data.

Anthropic's highest-capability Claude model. Current version: Opus 4.6. Excels at complex reasoning, extended analysis, nuanced writing, and tasks requiring deep domain knowledge. Higher cost and latency than Sonnet or Haiku. Use Opus for the tasks where quality is non-negotiable and throughput requirements are modest.

In a multi-agent system, the Claude agent responsible for decomposing goals, delegating subtasks to sub-agents, and synthesising results into a final output. The orchestrator maintains the high-level plan; sub-agents handle execution.

A unit of text generated by Claude in response to an API request. Output tokens are priced higher than input tokens. In extended thinking mode, thinking tokens (internal reasoning) are also billed as output tokens and priced accordingly.

In Claude Cowork, a distributable bundle that combines Skills, Connectors, and sub-agent configurations into a single installable package. Plugins allow enterprise teams to build and distribute specialised Cowork capabilities — a legal research plugin, a financial analysis plugin, a code review plugin — across their organisation.

A Claude API feature that caches a designated portion of the prompt (system prompt, static context, tool definitions) between API calls, charging a reduced rate for cache hits instead of reprocessing the full input. Can reduce API costs by 60–90% for applications with large, frequently reused system prompts.

An attack where malicious instructions are embedded in content Claude is asked to process — a document, a web page, a database record — intended to override Claude's system prompt or legitimate instructions. Enterprise Claude deployments should implement prompt injection defences, particularly in RAG and document processing pipelines.

A Claude.ai and Cowork feature that lets users maintain persistent context across multiple conversations. A Project stores documents, instructions, and conversation history so Claude has consistent knowledge about a topic, client, or workstream without needing to be re-briefed each session.

An architecture that combines a vector database with a language model. When a query is received, relevant documents are retrieved from the vector store and included in the Claude API request as context. RAG enables Claude to answer questions based on current, proprietary, or domain-specific knowledge that wasn't in its training data.

The maximum number of requests, tokens, or output tokens per minute/day allowed under a given Anthropic API tier. Rate limits vary by model and plan level. Enterprise teams typically negotiate higher limits. Exceeding a rate limit returns a 429 error and requires exponential backoff retry logic.

In Claude Code, a reusable, shareable unit of AI capability — a structured CLAUDE.md combined with tool configurations, MCP server connections, and instructions. Skills can be distributed across teams, enabling organisations to standardise how Claude behaves for specific tasks (code review, database query generation, test writing) without each developer configuring from scratch.

Anthropic's mid-tier Claude model, balancing capability and cost. Current version: Sonnet 4.6. The recommended default model for most enterprise applications — capable enough for complex tasks, priced for production scale. Used by the majority of enterprise Claude API deployments.

A Claude API response mode where tokens are delivered progressively as they are generated, rather than waiting for the full response. Streaming improves perceived responsiveness in interactive applications. Requires Server-Sent Events (SSE) handling on the client side.

A Claude instance delegated a specific subtask by an orchestrator agent. Sub-agents have isolated context, targeted tools, and a scoped objective. They operate independently and return results to the orchestrator. Sub-agents are the building block of scalable multi-agent systems in Claude Code and the Agent SDK.

An instruction block provided to Claude before the conversation begins. The system prompt defines Claude's role, persona, constraints, output format requirements, and available tools. In enterprise applications, the system prompt encodes business rules, compliance requirements, and domain-specific knowledge. It is the most important part of any Claude API integration.

The basic unit of text in Claude's processing. Approximately 1 token = 0.75 words in English. Claude's pricing, context limits, and processing are all measured in tokens. A typical A4 page of text is approximately 400–500 tokens.

The Claude API mechanism that allows Claude to invoke external functions during a conversation. You define tools as JSON schemas; Claude decides when to call them and with what parameters. Tool results are returned to Claude, which incorporates them into its response. The foundation of agentic Claude applications.

A database optimised for storing and searching high-dimensional vector embeddings. Used in RAG systems to store embedded documents and retrieve semantically relevant content at query time. Common enterprise choices: Pinecone, Weaviate, Qdrant, pgvector (PostgreSQL extension), Amazon OpenSearch.

Claude's capability to process and reason about images. Images can be passed via the Messages API as base64-encoded content or URLs. Enterprise use cases include document processing (scanned contracts, invoices, forms), diagram analysis, screenshot interpretation, and product image classification.

Administrative divisions within a Claude API account that allow different teams to have separate API keys, rate limit pools, usage tracking, and policy configurations. Useful for multi-team enterprise API deployments where different business units need isolated billing and governance.

Asking Claude to perform a task with no examples — relying on Claude's training and the clarity of the instruction alone. Works well for simple, well-defined tasks. For complex or structured tasks with specific format requirements, few-shot prompting typically outperforms zero-shot.